Deep Learning-based Image Analysis Research Articles

Abstract PO4-07-05: Clinical evaluation of image-based risk profiling in breast cancer histopathology and comparison to an established gene expression assay

Abstract Background: Among oestrogen receptor (ER)-positive and human epidermal growth factor receptor 2 (HER2)-negative early breast cancer, a significant proportion of patients are categorised as intermediate risk, based on classic clinico-pathological variables, thus providing limited information to guide adjuvant chemotherapy decisions. Prognostic risk profiling is an integrated part of modern breast cancer diagnostics to provide additional risk information for this patient group. Among the established prognostic assays based on gene expression, the Prosigna assay is widely used and provides an individual risk of recurrence (ROR) score and identifies intrinsic subtypes with associated survival outcomes. Pioneering artificial intelligence-based precision diagnostics, Stratipath Breast, is a deep learning-based image analysis tool that utilises digitised histopathological whole slide images to stratify intermediate risk patients in terms of risk of recurrence. Materials and methods: The study included 234 invasive breast tumours from patients with early ER-positive HER2-negative breast cancer, clinically assessed as intermediate risk tumours and eligible for chemotherapy. All tumours had therefore previously been analysed by the Prosigna assay in clinical routine at point of diagnosis between 2020 and 2022 at the Karolinska University Hospital and Södersjukhuset, Stockholm, Sweden. Clinicopathological data including Prosigna results (ROR score, risk group and intrinsic subtype) were extracted from medical records, along with the corresponding archived haematoxylin and eosin (HE)-stained formalin-fixed paraffin-embedded tissue slides. The HE slides were subsequently digitised and analysed by the Stratipath Breast tool. The agreement between the two tests for risk stratification was evaluated in this real-world breast cancer case series. Results: The Prosigna assay classified 76 (32.5%), 110 (47.0%) and 48 (20.5%) tumours as low, intermediate and high risk, respectively. The Stratipath Breast analysis identified 116 (49.6%) tumours as low risk and 118 (50.4%) as high risk. Among Prosigna intermediate risk tumours, 52 (47.3%) were stratified as low risk and 58 (52.7%) as high risk by Stratipath Breast. The overall agreement between the two tests for low risk and high risk groups was 71.0%, with a Cohen’s linear kappa of 0.42. Twelve of the 48 Prosigna high risk cases were classified as Stratipath low risk. ROR scores were higher in the Stratipath high risk group compared to the low risk group (p < 0.001), across all cases as well as in the Prosigna intermediate group. Among the 176 histological grade (NHG) 2 tumours, 97 (55.1%) and 79 (44.9%) were stratified as Stratipath low risk and high risk, respectively, whereas 66 (37.5%), 83 (47.2%) and 27 (15.3%) were stratified as Prosigna low, intermediate and high risk, respectively. The majority of NHG1 (10 of 12) and NHG3 (37 of 46) tumours were stratified as Stratipath low risk and high risk, respectively. For both risk profiling tests, NHG and Ki67 proliferation index differed between risk groups. Conclusions: In this study of clinically assessed intermediate risk ER-positive HER2-negative breast cancer, we observed a moderate agreement between Prosigna and Stratipath Breast for low risk and high risk groups. In addition, image-based risk profiling stratified more of the NHG2 tumours as high risk. Citation Format: Stephanie Robertson, Yinxi Wang, Wenwen Sun, Emelie Karlsson, Sandy Kang Lövgren, Mattias Rantalainen, Johan Hartman. Clinical evaluation of image-based risk profiling in breast cancer histopathology and comparison to an established gene expression assay [abstract]. In: Proceedings of the 2023 San Antonio Breast Cancer Symposium; 2023 Dec 5-9; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2024;84(9 Suppl):Abstract nr PO4-07-05.

Abstract 2319: Deep learning-enabled dynamic infiltration and response to NK therapies in solid tumor organoids

Abstract Patient-derived organoid co-culture models have become a state of the art platform to evaluate the effectiveness of immunotherapeutic agents in preclinical studies. However, preclinical development as well as quantification of therapeutic potency and resistance require differential labeling of cells with fluorescent dyes or transgenic fluorescent proteins, which may dysregulate the effector cell function or produce undesired cytotoxicity. Deep learning-based image analysis allows the use of brightfield images to co-localize tumor organoids (TOs) with effector cells, and measure TO-specific responses to novel candidate immunotherapies in a label-free manner; this can accelerate their evaluation as therapeutic candidates in cancer patients and shed light on tumor-immune interaction mechanisms. Here we record multi-day time-lapse confocal microscopy images of TOs co-cultured with NK cells at increasing concentrations of target to effector cells. Two pre-trained U-Net convolutional neural networks are used to separately segment tumor and immune cells from the brightfield channel, and quantify immune cell infiltration over time. TO segmentation masks are registered with a vital dye (caspase 3/7) channel to selectively quantify TO cell death and correlate it with estimated immune cell infiltration across a large cohort of patient-derived TO models (including breast, colorectal, endometrial, gastric, head and neck, liver, lung and pancreatic cancer), at increasing concentrations of 6 different NK cell types (identified as A through F). We find that the estimated density of infiltrating immune cells is highly correlated with TO death, as quantified by fluorescence intensity of caspase over time (median Pearson’s r=0.89, P<0.001). Higher ratios of effector to target cells lead to higher activation of NK cells as measured by infiltrating cell density. Differential infiltration dynamics are observed across TOs and immune cell lines, as peak infiltration density is affected by co-culture time, ratio of target to effector cells, TO line, and NK cell type. These findings highlight that the paired segmentation models are able to measure varying degrees of infiltration across different types of immune cells and TO models. The approach described here is highly scalable and only requires capturing brightfield images of the assay, therefore eliminating the need to label cells and avoiding phototoxic effects or alteration of the immune cells activity. Time series measurements enable quantification of patterns of immune cell activation, including infiltration, migration and co-localization dynamics, which provide insights into the pharmacokinetics and mechanisms for specific immune therapies. Overall, this methodology enables high throughput screening of many therapeutic candidates across dozens to hundreds of unique TO-models, thereby facilitating targeted precision therapy. Citation Format: Luca Lonini, Sonal Khare, Timothy Don Lopez, Stanislaw Szydlo, Michael Streit, Olga A. Karginova, Mary K. Flaherty, Ameen Salhaudeen, Andre Kunert, Anna Oja, Kate Sasser, Martin C. Stumpe, Chi-Sing Ho. Deep learning-enabled dynamic infiltration and response to NK therapies in solid tumor organoids [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 2319.

Abstract 3650: Deep learning-based image cytometry and co-localization index in tumor immune microenvironment

Abstract Background: In pathology, digitizing tissue slides has prompted a remarkable development in image analysis using deep learning. This technological advancement is anticipated to aid in pathological diagnosis and to enhance patient management. Deep learning-based image cytometry (DL-IC) enables accurate cell identification and counting and the acquisition of vast amounts of location information from tissue slides. DL-IC can capture information about the diverse and complex tumor immune microenvironment(TIME) and its constituent cells and help identify biomarkers to predict patient treatment efficacy and prognosis. This study will introduce a spatial interaction map and co-localization index (CLI) for the analysis of TIME using DL-IC. Materials and Methods: Cu-Cyto, a deep learning-based image analysis technology, was used in this study; bit-pattern kernel filtering technology, which can accurately count cells while avoiding the determination of multiple cell counts, was used by Cu-Cyto (Abe T, et al. Anticancer Res. 43:3755, 2023). First, the accuracy of cell counting using Cu-Cyto was evaluated. Second, tumor tissue slides with immunohistochemical (IHC) and hematoxylin-eosin (H&E) staining were prepared from surgical specimens of patients with rectal cancer who had undergone neoadjuvant chemoradiotherapy (NACRT), and the relationship between the co-localization index (CLI) of cancer cells and CD8+T cells and prognosis was investigated. CLI was defined to predict cell- cell interactions on the basis of the relative distances between different cell types (Nagasaka T. PCT/JP 2021/021455). Results: The performances of three versions of Cu-Cyto were evaluated according to their learning stages. In the early stage of learning, the F1 score for immunostained CD8+ T cells (0.343) was higher than that for non-immunostained cells (adenocarcinoma cells [0.040] and lymphocytes [0.002]). In the latest stage of learning, the F1 scores for adenocarcinoma cells, lymphocytes, and CD8+ T cells were 0.589, 0.889, and 0.911, respectively. Next, we examined the correlation of CLI between cancer cells and CD8+ T cells with prognosis: patients with a higher CLI significantly prolonged five-year disease-free survival (P=0.038), while there was no substantial difference in five-year overall survival (P=0.57). Conclusion]: Cu-Cyto performed well in cell determination. In particular, IHC was able to increase the learning efficiencies in the early stages of learning. The CLI calculated using Cu-Cyto ts an objective, reproducible, and innovative quantitative approach for assessing cell-cell interactions, which has been shown to be associated with recurrence-free survival in patients with rectal cancer after NACRT. Its performance is expected to improve even further with continuous learning, and the DL-IC can contribute to the implementation of precision oncology. Citation Format: Tomoki Abe, Kimihiro Yamashita, Toru Nagasaka, Tomosuke Mukoyama, Souichirou Miyake, Yasuhiro Ueda, Masayuki Ando, Yuki Okazoe, Takao Tsuneki, Yukari Adachi, Ryunosuke Konaka, Ryuichiro Sawada, Hironobu Goto, Hiroshi Hasegawa, Shingo Kanaji, Takeru Matsuda, Takumi Fukumoto, Yoshihiro Kakeji. Deep learning-based image cytometry and co-localization index in tumor immune microenvironment [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 3650.

Analysis of Osteosarcoma Cell Lines and Patient Tissue Using a 3D In Vivo Tumor Model—Possible Effects of Punicalagin

Osteosarcomas are the most common primary malignant bone tumors and mostly affect children, adolescents, and young adults. Despite current treatment options such as surgery and polychemotherapy, the survival of patients with metastatic disease remains poor. In recent studies, punicalagin has reduced the cell viability, angiogenesis, and invasion in cell culture trials. The aim of this study was to examine the effects of punicalagin on osteosarcomas in a 3D in vivo tumor model. Human osteosarcoma biopsies and SaOs-2 and MG-63 cells, were grown in a 3D in vivo chorioallantoic membrane (CAM) model. After a cultivation period of up to 72 h, the tumors received daily treatment with punicalagin for 4 days. Weight measurements of the CAM tumors were performed, and laser speckle contrast imaging (LSCI) and a deep learning-based image analysis software (CAM Assay Application v.3.1.0) were used to measure angiogenesis. HE, Ki-67, and Caspase-3 staining was performed after explantation. The osteosarcoma cell lines SaOs-2 and MG-63 and osteosarcoma patient tissue displayed satisfactory growth patterns on the CAM. Treatment with punicalagin decreased tumor weight, proliferation, and tumor-induced angiogenesis, and the tumor tissue showed pro-apoptotic characteristics. These results provide a robust foundation for the implementation of further studies and show that punicalagin offers a promising supplementary treatment option for osteosarcoma patients. The 3D in vivo tumor model represents a beneficial model for the testing of anti-cancer therapies.

Direct Observation and Automated Measurement of Stomatal Responses to Pseudomonas syringae pv. tomato DC3000 in Arabidopsis thaliana.

Stomata are microscopic pores found in the plant leaf epidermis. Regulation of stomatal aperture is pivotal not only for balancing carbon dioxide uptake for photosynthesis and transpirational water loss but also for restricting bacterial invasion. While plants close stomata upon recognition of microbes, pathogenic bacteria, such as Pseudomonas syringae pv. tomato DC3000 (Pto), reopen the closed stomata to gain access into the leaf interior. In conventional assays for assessing stomatal responses to bacterial invasion, leaf epidermal peels, leaf discs, or detached leaves are floated on bacterial suspension, and then stomata are observed under a microscope followed by manual measurement of stomatal aperture. However, these assays are cumbersome and may not reflect stomatal responses to natural bacterial invasion in a leaf attached to the plant. Recently, a portable imaging device was developed that can observe stomata by pinching a leaf without detaching it from the plant, together with a deep learning-based image analysis pipeline designed to automatically measure stomatal aperture from leaf images captured by the device. Here, building on these technical advances, a new method to assess stomatal responses to bacterial invasion in Arabidopsis thaliana is introduced. This method consists of three simple steps: spray inoculation of Pto mimicking natural infection processes, direct observation of stomata on a leaf of the Pto-inoculated plant using the portable imaging device, and automated measurement of stomatal aperture by the image analysis pipeline. This method was successfully used to demonstrate stomatal closure and reopening during Pto invasion under conditions that closely mimic the natural plant-bacteria interaction.

Applying traffic camera and deep learning-based image analysis to predict PM2.5 concentrations

BackgroundAir pollution has caused a significant burden in terms of mortality and mobility worldwide. However, the current coverage of air quality monitoring networks is still limited. ObjectiveThis study aims to apply a novel approach to convert the existing traffic cameras into sensors measuring particulate matter with a diameter of 2.5 μm or less (PM2.5) so that the coverage of PM2.5 monitoring could be expanded without extra cost. MethodsIn our study, the traffic camera images were collected at a rate of 4 images/h and the corresponding hourly PM2.5 concentration was collected from the reference grade PM2.5 station 3 km away. A customized neural network model was trained to obtain the PM2.5 concentration from images followed by a random forest model to predict the hourly PM2.5 concentration. The saliency maps and the feature importance were utilized to interpret the neural network. ResultsProposed novel approach has a high prediction performance to predict hourly PM2.5 from traffic camera images, with a root mean square error (RMSE) of 0.76 μg/m3 and a coefficient of determination (R2) of 0.98. The saliency map shows neural network focuses on unobstructed far-end road surfaces while the random forest feature importance highlights the first quarter image's significance. The model performance is robust whether weather conditions are controlled or not. ConclusionOur study provided a practical approach to converting the existing traffic cameras into PM2.5 sensors. The deep learning method based on the Resnet architecture in our study can broaden the coverage of PM2.5 monitoring with no additional infrastructure needed.

Deep learning-based image analysis for in situ microscopic imaging of cell culture process

Mammalian cell culture is an important bioprocess that directly affects the quality and yield of biopharmaceuticals. Traditionally, condition monitoring of the operation is based on sampling periodically and off-line analysis, which is labor intensive, time consuming, and causing time delays. In this work, in situ microscope is investigated for on-line real-time monitoring of the culture process of Chinese hamster ovary cells with focus on investigation of deep learning-based Mask R-CNN algorithm for image analysis. The model is trained by 184 images with 183,040 cells using data augmentation methods and transfer learning technique. Mask R-CNN segmented the clustered cells more effectively than the conventional one combining edge detection, intensity thresholding, and advanced watershed method as well as the multi-scale edge detection method. Its Dice score, accuracy, precision, sensitivity, F1 score, specificity, and relative volume difference reach 0.862, 0.945, 0.901, 0.827, 0.862, 0.977, and 0.082, respectively. The evolution of geometrical features of cells were further analyzed, including equivalent diameter, circularity, aspect ratio, and eccentricity. The result demonstrated the great potential of deep learning technology in analysis of on-line images for optimization and control of the cell culture process.

Upregulation of TIM3 and reduced expression of PD-1 on immune cell subsets in advanced prostate cancers

Abstract Introduction/Objective Although most prostate cancers behave in an indolent manner, a small proportion is highly aggressive. Both primary and advanced prostate cancer is widely known as a non-inflamed cancer that is characterized by a paucity of immune infiltration. Methods/Case Report To assess the spatial interplay of more than 30 TIM3, CTLA-4, PD-1/-L1 expressing leukocyte subpopulations in 453 prostate cancers, tissue microarrays were stained with 21 antibodies using our BLEACH&STAIN multiplex fluorescence immunohistochemistry approach and analyzed using a deep learning-based image analysis framework. Results (if a Case Study enter NA) The immune cell density of CD8+ cytotoxic T-cells, CD4+ T-helper cells, FOXP3+ regulatory T-cells, M1/ M2 macrophages, as well as CD11c+ dendritic cells increased consistently along with the Gleason grade in primary prostate cancer (p≤ 0.034 each). In recurrent prostate cancers under therapy, the density of FOXP3+regulatory T-cells and M1 macrophages further increased, while the density of CD8+ cytotoxic T-cells, CD4+ T-helper cells, as well as CD11c+ dendritic cells decreased (p≤0.017 each). Although the immune checkpoint expression of TIM3 on T-cell subsets, macrophages and dendritic cells was upregulated in advanced/ recurrent tumors, the expression level of PD-1 was downregulated in all analyzed T-cell subsets. Conclusion Although prostate cancer is a generally considered a low inflamed tumor, the degree of tumor infiltration by various immune cell subtypes increases markedly along with tumor progression and in recurrent tumors under therapy. Taken together, these data suggest that the evaluation of the spatial distribution of immune cell types along with their immune checkpoint expression can provide relevant clinical information in prostate cancer.

Deep learning-based image analysis identifies a DAT-negative subpopulation of dopaminergic neurons in the lateral Substantia nigra

Here we present a deep learning-based image analysis platform (DLAP), tailored to autonomously quantify cell numbers, and fluorescence signals within cellular compartments, derived from RNAscope or immunohistochemistry. We utilised DLAP to analyse subtypes of tyrosine hydroxylase (TH)-positive dopaminergic midbrain neurons in mouse and human brain-sections. These neurons modulate complex behaviour, and are differentially affected in Parkinson’s and other diseases. DLAP allows the analysis of large cell numbers, and facilitates the identification of small cellular subpopulations. Using DLAP, we identified a small subpopulation of TH-positive neurons (~5%), mainly located in the very lateral Substantia nigra (SN), that was immunofluorescence-negative for the plasmalemmal dopamine transporter (DAT), with ~40% smaller cell bodies. These neurons were negative for aldehyde dehydrogenase 1A1, with a lower co-expression rate for dopamine-D2-autoreceptors, but a ~7-fold higher likelihood of calbindin-d28k co-expression (~70%). These results have important implications, as DAT is crucial for dopamine signalling, and is commonly used as a marker for dopaminergic SN neurons.

Assessment of deep learning-based image analysis for disaster waste identification

Accurately predicting different types and quantities of wastes is crucial for proper waste management and sustainable growth and development. High levels of disaster waste (DW) can negatively impact emergency responses, public health, and recovery and rebuilding processes, making the effective management of DW highly important. Most existing waste identification methods cannot suitably identify DW since it is highly subjective in form. In this study, we examined the applicability of different deep learning-based image recognition techniques for DW recognition. A DW image dataset was collected and categorized into normal and hard cases. Then, experiments were conducted using these three network models: DeepLabV3+, ResNeSt-50, and ResNeSt-101. The experimental results indicated that the mean intersection over union (mIoU) of the DeepLabV3+ model was 0.802 for the normal case and that of the ResNeSt-101 model was 0.676 for the hard case, indicating favorable performances. Overall, the deep learning-based image analysis method is more robust than existing methods, particularly in terms of its ability to accurately identify waste in various forms, such as mixed or piled up waste.

Deep learning-assisted high-content screening identifies isoliquiritigenin as an inhibitor of DNA double-strand breaks for preventing doxorubicin-induced cardiotoxicity

BackgroundAnthracyclines including doxorubicin are essential components of many cancer chemotherapy regimens, but their cardiotoxicity severely limits their use. New strategies for treating anthracycline-induced cardiotoxicity (AIC) are still needed. Anthracycline-induced DNA double-strand break (DSB) is the major cause of its cardiotoxicity. However, DSB-based drug screening for AIC has not been performed possibly due to the limited throughput of common assays for detecting DSB. To discover new therapeutic candidates for AIC, here we established a method to rapidly visualize and accurately evaluate the intranuclear anthracycline-induced DSB, and performed a screening for DSB inhibitors.ResultsFirst, we constructed a cardiomyocyte cell line stably expressing EGFP-53BP1, in which the formation of EGFP-53BP1 foci faithfully marked the doxorubicin-induced DSB, providing a faster and visible approach to detecting DSB. To quantify the DSB, we used a deep learning-based image analysis method, which showed the better ability to distinguish different cell populations undergoing different treatments of doxorubicin or reference compounds, compared with the traditional threshold-based method. Subsequently, we applied the deep learning-assisted high-content screening method to 315 compounds and found three compounds (kaempferol, kaempferide, and isoliquiritigenin) that exert cardioprotective effects in vitro. Among them, the protective effect of isoliquiritigenin is accompanied by the up-regulation of HO-1, down-regulation of peroxynitrite and topo II, and the alleviation of doxorubicin-induced DSB and apoptosis. The results of animal experiments also showed that isoliquiritigenin maintained the myocardial tissue structure and cardiac function in vivo. Moreover, isoliquiritigenin did not affect the killing of HeLa and MDA-MB-436 cancer cells by doxorubicin and thus has the potential to be a lead compound to exert cardioprotective effects without affecting the antitumor effect of doxorubicin.ConclusionsOur findings provided a new method for the drug discovery for AIC, which combines phenotypic screening with artificial intelligence. The results suggested that isoliquiritigenin as an inhibitor of DSB may be a promising drug candidate for AIC.

Prediction of lymph node metastasis in primary gastric cancer from pathological images and clinical data by multimodal multiscale deep learning

Background:Gastric cancer is the third most common cause of cancer-related death. Accurate preoperative prediction of lymph node metastasis (LNM) in primary gastric cancer strongly influences the choice of surgical approach and the prognosis of gastric cancer patients. Objective:To develop and validate a deep learning-based model to analyze routine histological slides of patients with primary gastric cancer as well as clinical data to predict the occurrence of LNM preoperatively. Patients and Methods:Radical surgery slides from 309 patients and biopsy slides from 157 patients were collected from Fujian Cancer Hospital, along with radical surgery slides from 306 patients from The Cancer Genome Atlas (TCGA). Clinical data, including age, gender, lauren classification, and tumor location, were collected. These datasets were used to develop and validate a deep learning-based model. Results:Five models were trained via cross-validation, with a mean area under the receiver operating characteristic curve (AUC) (standard deviation [SD]) of 0.877 (0.048) achieved. There was a significant difference in scores between both classes (LNM positive [N+] and LNM negative [N0]) ( p<0.001). we validated the performance of the model on biopsy slides and achieved a mean AUC (SD) of 0.725 (0.020). In the analysis of clinical data, the lauren classification was demonstrated to be an independent risk factor for predicting LNM. Conclusion:Our study confirmed that deep learning-based image analysis could preoperatively predict LNM in patients with primary gastric cancer, combining histological slides at different magnification scales and relevant clinical data, showing superiority over individual modality prediction.

A Review on Imaging Techniques and Artificial Intelligence Models for Osteoporosis Prediction.

Osteoporosis causes harmful influences on both men and women of all races. Bone mass, also referred to as "bone density," is frequently used to assess the health of bone. Humans frequently experience bone fractures as a result of trauma, accidents, metabolic bone diseases, and disorders of bone strength, which are typically led by changes in mineral composition and result in conditions like osteoporosis, osteoarthritis, osteopenia, etc. Artificial intelligence holds a lot of promise for the healthcare system. Data collection and preprocessing seem to be more essential for analysis, so bone images from different modalities, such as X-ray, Computed Tomography (CT), and Magnetic Resonance Imaging (MRI), are taken into consideration that help to recognize, classify, and evaluate the patterns in clinical images. This research presents a comprehensive overview of the performance of various image processing techniques and deep learning approaches used to predict osteoporosis through image segmentation, classification, and fault detection. This survey outlined the proposed domain-based deep learning model for image classification in addition to the initial findings. The outcome identifies the flaws in the existing literature's methodology and lays the way for future work in the deep learning-based image analysis model.

Enhancing Photovoltaic Module Fault Diagnosis with Unmanned Aerial Vehicles and Deep Learning-Based Image Analysis

Artificial intelligence (AI) has evolved into a powerful tool that has wide-spread application in computer vision such as computer-aided inspection, industrial control systems, and navigation of robots. Monitoring the condition of machineries and mechanical components for the presence of faults with the aid of image-based automated analysis is one major application of computer vision. Diagnosing machinery faults from images can be made feasible with the adoption of deep learning and machine learning techniques. The primary objective of this study is to detect malfunctions in photovoltaic (PV) modules by utilizing a combination of deep learning and machine learning methodologies, with the assistance of RGB images captured via unmanned aerial vehicles. Six test conditions of PV modules such as good panel, snail trail, delamination, glass breakage, discoloration, and burn marks were considered in the study. The overall experimentation was carried out in two phases: (i) deep learning phase and (ii) machine learning phase. In the initial deep learning phase, the final fully connected layer of six pretrained networks, namely, DenseNet-201, VGG19, ResNet-50, GoogLeNet, VGG16, and AlexNet, was utilized to extract PVM image features. During the machine learning phase, feature selection from the extracted features was carried out using the J48 decision tree algorithm. Post selection of features, three families of classifiers such as tree, Bayes, and lazy were applied to determine the best feature extractor-classifier pair. The combination of DenseNet-201 features with k-nearest neighbour (IBK) classifier produced the overall classification accuracy of 100.00% among all other pretrained network features and classifiers considered.

Development of a Deep Learning Model for Retinal Hemorrhage Detection on Head Computed Tomography in Young Children

Abusive head trauma (AHT) in children is often missed in medical encounters, and retinal hemorrhage (RH) is considered strong evidence for AHT. Although head computed tomography (CT) is obtained routinely, all but exceptionally large RHs are undetectable on CT images in children. To examine whether deep learning-based image analysis can detect RH on pediatric head CT. This diagnostic study included 301 patients diagnosed with AHT who underwent head CT and dilated fundoscopic examinations at a quaternary care children's hospital. The study assessed a deep learning model using axial slices from 218 segmented globes with RH and 384 globes without RH between May 1, 2007, and March 31, 2021. Two additional light gradient boosting machine (GBM) models were assessed: one that used demographic characteristics and common brain findings in AHT and another that combined the deep learning model's risk prediction plus the same demographic characteristics and brain findings. Sensitivity (recall), specificity, precision, accuracy, F1 score, and area under the curve (AUC) for each model predicting the presence or absence of RH in globes were assessed. Globe regions that influenced the deep learning model predictions were visualized in saliency maps. The contributions of demographic and standard CT features were assessed by Shapley additive explanation. The final study population included 301 patients (187 [62.1%] male; median [range] age, 4.6 [0.1-35.8] months). A total of 120 patients (39.9%) had RH on fundoscopic examinations. The deep learning model performed as follows: sensitivity, 79.6%; specificity, 79.2%; positive predictive value (precision), 68.6%; negative predictive value, 87.1%; accuracy, 79.3%; F1 score, 73.7%; and AUC, 0.83 (95% CI, 0.75-0.91). The AUCs were 0.80 (95% CI, 0.69-0.91) for the general light GBM model and 0.86 (95% CI, 0.79-0.93) for the combined light GBM model. Sensitivities of all models were similar, whereas the specificities of the deep learning and combined light GBM models were higher than those of the light GBM model. The findings of this diagnostic study indicate that a deep learning-based image analysis of globes on pediatric head CTs can predict the presence of RH. After prospective external validation, a deep learning model incorporated into CT image analysis software could calibrate clinical suspicion for AHT and provide decision support for which patients urgently need fundoscopic examinations.

Deep Learning–Based Image Analysis of Liver Steatosis in Mouse Models

The incidence of non-alcoholic fatty liver disease (NAFLD) is a continuously growing health problem worldwide, along with obesity. Therefore, both novel methods to efficiently study the manifestation of NAFLD and to analyze drug efficacy in pre-clinical models are needed. In the present study, we developed a deep neural network -based model to quantify micro- and macrovesicular steatosis in the liver on hematoxylin-eosin stained whole slide images (WSIs), using the cloud-based platform, Aiforia Create (Aiforia Technologies, Helsinki, Finland). The training data included a total of 101 WSIs from dietary interventions of wild-type mice and from two genetically modified (GM) mouse models with steatosis. The algorithm was trained for the following: to detect liver parenchyma, to exclude the blood vessels and any artefacts generated during tissue processing and image acquisition, to recognize and differentiate the areas of micro- and macrovesicular steatosis, and to quantify the recognized tissue area. The results of the image analysis replicated well the evaluation by expert pathologists, and correlated well with the liver fat content measured by EcoMRI ex vivo, and the correlation with total liver triglycerides were notable. In conclusion, the developed deep learning-based model is a novel tool for studying liver steatosis in mouse models on paraffin sections, and thus, can facilitate reliable quantification of the amount of steatosis in large preclinical study cohorts.

Label free identification of different cancer cells using deep learning-based image analysis

Cancer diagnostics is an important field of cancer recovery and survival with many expensive procedures needed to administer the correct treatment. Machine Learning (ML) approaches can help with the diagnostic prediction from circulating tumor cells in liquid biopsy or from a primary tumor in solid biopsy. After predicting the metastatic potential from a deep learning model, doctors in a clinical setting can administer a safe and correct treatment for a specific patient. This paper investigates the use of deep convolutional neural networks for predicting a specific cancer cell line as a tool for label free identification. Specifically, deep learning strategies for weight initialization and performance metrics are described, with transfer learning and the accuracy metric utilized in this work. The equipment used for prediction involves brightfield microscopy without the use of chemical labels, advanced instruments, or time-consuming biological techniques, giving an advantage over current diagnostic methods. In the procedure, three different binary datasets of well-known cancer cell lines were collected, each having a difference in metastatic potential. Two different classification models were adopted (EfficientNetV2 and ResNet-50) with the analysis given for each stage in the ML architecture. The training results for each model and dataset are provided and systematically compared. We found that the test set accuracy showed favorable performance for both ML models with EfficientNetV2 accuracy reaching up to 99%. These test results allowed EfficientNetV2 to outperform ResNet-50 at an average percent increase of 3.5% for each dataset. The high accuracy obtained from the predictions demonstrates that the system can be retrained on a large-scale clinical dataset.

In vitro cardiac modeling of desmin-related myofibrillar myopathy and deep learning-based image analysis to develop a high-content screening assay

Mutations of DES, the gene encoding desmin, the main intermediate filament of muscle cells, are one of the major cause of Myofibrillar myopathy(MFM). This disease is mainly characterized by a muscular dystrophy but also, in many cases, by the progressive emergence of cardiac dysfunctions evolving frequently from dilated cardiomyopathy to heart failure. For the moment, no specific treatment can revert this disease. The following study relied on the development of an automated tool base on deep-learning to discriminate efficiently mutant and control cardiomyocytes in the perspective of drug screening. We use cardiomyocytes derived from human induced pluripotent stem cells (hiPSC-CMs) obtained from control or patients carrying a DES mutation (E439K). Phenotyping of hiPSC-CMs was performed by immunocytostaining of the sarcomeres or the desmin network. After imaging, the morphological perturbations were evaluated manually based either on the sarcomeric organization or on the aspect of the desmin network at single cell level. Then an automated classification model was generated by using machine-learning method trained with the previous images annotated with the results of the manual classifications. In parallel, another automated classification model based on cell profiling from sarcomeric, desmin and mitochondrial staining by using the deep learning EfficientNet B0 model was setting up in order to distinguish Control cell and DES mutated cell. Manual classification based on sarcomeric organization was established on 4 classes ranging from cells with disorganized sarcomeres to cells with a well-organized sarcomeric network on 1500 single cells. We observed that the distribution of DES mutated cells is shifted toward the class of disorganized sarcomere compared to control cells. Identically, we also manually classified 1200 single cells based on the organization of the desmin network (4 classes). We also observed a shift of DES mutant cells to be classified as cells with aggregated desmin network compared to control cells. Finally, the various validation tests of the two automated classification methods showed significant correlation with the manual classification methods. Here we demonstrate the efficacy of our automated classification software to be able to discriminates mutant and control cells. This analysis platform will then be used to perform high throughput drug screening assays to discover new therapeutic molecules.

A convolutional attention mapping deep neural network for classification and localization of cardiomegaly on chest X-rays

Building a reliable and precise model for disease classification and identifying abnormal sites can provide physicians assistance in their decision-making process. Deep learning based image analysis is a promising technique for enriching the decision making process, and accordingly strengthening patient care. This work presents a convolutional attention mapping deep learning model, Cardio-XAttentionNet, to classify and localize cardiomegaly effectively. We revisit the global average pooling (GAP) system and add a weighting term to develop a light and effective Attention Mapping Mechanism (AMM). The model enables the classification of cardiomegaly from chest X-rays through image-level classification and pixel-level localization only from image-level labels. We leverage some of the advanced ConvNet architectures as a backbone-model of the proposed attention mapping network to build Cardio-XAttentionNet. The proposed model is trained on ChestX-Ray14, which is a publicly accessible chest X-ray dataset. The best single model achieves an overall precision, recall, F-1 measure and area under curve (AUC) scores of 0.87, 0.85, 0.86 and 0.89, respectively, for the classification of the cardiomegaly. The results also demonstrate that the Cardio-XAttentionNet model well captures the cardiomegaly class information at image-level as well as localization at pixel-level on chest x-rays. A comparative analysis between the proposed AMM and existing GAP based models shows that the proposed model achieves a state-of-the-art performance on this dataset for cardiomegaly detection using a single model.

Abstract LB025: First disclosure of AZD5335, a TOP1i-ADC targeting low and high FRα-expressing ovarian cancer with superior preclinical activity vs FRα-MTI ADC

Abstract Folate receptor alpha (FRα) is a cell surface GPI-anchored protein overexpressed in several solid tumors with highest prevalence in ovarian cancer and lung adenocarcinoma but restricted expression in normal tissues. An antibody drug conjugate (ADC) with a microtubule inhibitor (MTI) payload recently received accelerated approval from the FDA for FRα-expressing platinum-resistant ovarian cancer. We describe for the first time the preclinical activity of AZD5335, an FRα-targeting antibody conjugated to AZ’s proprietary topoisomerase 1 inhibitor (TOP1i) payload, AZ14170132, with a homogeneous drug-to-antibody ratio of 8 (DAR8) and potential benefits vs an MTI-based ADC. AZD5335’s primary mechanism of action is to deliver TOP1i payload into FRα-expressing cancer cells, leading to DNA damage and apoptotic cell death. The TOP1i payload mediates bystander killing, which is important for targeting tumors with less than uniformly positive expression. Here, we report that a single dose of AZD5335 at 2.5 mg/kg was sufficient to provide a robust and durable anti-tumor response in FRα-expressing ovarian cancer cell line xenografts (CDX) with a tumor growth inhibition (TGI) of 75%-94% and median best tumor volume reduction &amp;gt;30% in 14/17 (82%) ovarian cancer patient-derived xenografts (PDX) evaluated. FRα-expression levels (by IHC and deep-learning based image analysis) correlated with efficacy in the tested PDX models, and we observed that AZD5335 was also active in models with low levels of target expression (75% of cells with FRα staining of 2+), expected to be representative of patients who would be ineligible for treatment with the MTI-ADC. Furthermore, AZD5335 demonstrated superior activity vs an FRα-MTI benchmark ADC with respect to anti-tumor activity and duration of response in two PDX models with low-to-medium FRα expression at equal or higher drug doses (e.g., in OV0857-CIS: 96% TGI vs 24% TGI at 5 mg/kg and 95% TGI vs 2% TGI at 2.5 mg/kg of a single IV dose AZD5335 and FRα-MTI, respectively). These data indicate that AZD5335 is a promising therapeutic candidate for the treatment of ovarian cancers across the spectrum of FRα-expression. Citation Format: Marco Gymnopoulos, Tima Thomas, Diana Gasper, Judith Anderton, Ravinder Tammali, Ed Rosfjord, Nick Durham, Chris Ward, Claire Myers, Jixin Wang, Wenyan Zhong, Simon Christ, Lina Meinecke, Katharina Nekolla, Laura Sebastian Monasor, Roger Dodd, Neki Patel, Mark Albertella, Jorge Zeron-Medina, Paula Fraenkel, Puja Sapra. First disclosure of AZD5335, a TOP1i-ADC targeting low and high FRα-expressing ovarian cancer with superior preclinical activity vs FRα-MTI ADC [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 2 (Clinical Trials and Late-Breaking Research); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(8_Suppl):Abstract nr LB025.