Patrol Security Game: Defending against Adversary with Freedom in Attack Timing, Location, and Duration

Deep learning has become increasingly popular due to its remarkable ability to learn high-dimensional feature representations. Numerous algorithms and models have been developed to enhance the application of deep learning across various real-world tasks, including image classification, natural language processing, and autonomous driving. However, deep learning models are susceptible to backdoor threats, where an attacker manipulates the training process or data to cause incorrect predictions on malicious samples containing specific triggers, while maintaining normal performance on benign samples. With the advancement of deep learning, including evolving training schemes and the need for large-scale training data, new threats in the backdoor domain continue to emerge. Conversely, backdoors can also be leveraged to protect deep learning models, such as through watermarking techniques. In this thesis, we conduct an in-depth investigation into backdoor techniques from three novel perspectives. In the first part of this thesis, we demonstrate that emerging deep learning training schemes can introduce new backdoor risks. Specifically, pre-trained Natural Language Processing (NLP) models can be easily adapted to a variety of downstream language tasks, significantly accelerating the development of language models. However, the pre-trained model becomes a single point of failure for these downstream models. We propose a novel task-agnostic backdoor attack against pre-trained NLP models, wherein the adversary does not need prior information about the downstream tasks when implanting the backdoor into the pre-trained model. Any downstream models transferred from this malicious model will inherit the backdoor, even after extensive transfer learning, revealing the severe vulnerability of pre-trained foundation models to backdoor attacks. In the second part of this thesis, we develop novel backdoor attack methods suited to new threat scenarios. The rapid expansion of deep learning models necessitates large-scale training data, much of which is unlabeled and outsourced to third parties for annotation. To ensure data security, most datasets are read-only for training samples, preventing the addition of input triggers. Consequently, attackers can only achieve data poisoning by uploading malicious annotations. In this practical scenario, all existing data poisoning methods that add triggers to the input are infeasible. Therefore, we propose new backdoor attack methods that involve poisoning only the labels without modifying any input samples. In the third part of this thesis, we utilize the backdoor technique to proactively protect our deep learning models, specifically for intellectual property protection. Considering the complexity of deep learning tasks, generating a well-trained deep learning model requires substantial computational resources, training data, and expertise. Therefore, it is essential to protect these assets and prevent copyright infringement. Inspired by backdoor attacks that can induce specific behaviors in target models through carefully designed samples, several watermarking methods have been proposed to protect the intellectual property of deep learning models. Model owners can train their models to produce unique outputs for certain crafted samples and use these samples for ownership verification. While various extraction techniques have been designed for supervised deep learning models, challenges arise when applying them to deep reinforcement learning models due to differences in model features and scenarios. Therefore, we propose a novel watermarking scheme to protect deep reinforcement learning models from unauthorized distribution. Instead of using spatial watermarks as in conventional deep learning models, we design temporal watermarks that minimize potential impact and damage to the protected deep reinforcement learning model while achieving high-fidelity ownership verification. In summary, this thesis investigates the evolving landscape of backdoor threats during the development of deep learning techniques and the use of backdoors for beneficial purposes in intellectual property protection.

Purpose: Although deep learning (DL) models have shown increasing ability to accurately classify diagnostic images in oncology, significantly large amounts of well-curated data are often needed to match human level performance. Given the relative paucity of imaging datasets for less prevalent cancer types, there is an increasing need for methods which can improve the performance of deep learning models trained using limited diagnostic images. Deep metric learning (DML) is a potential method which can improve accuracy in deep learning models trained on limited datasets. Leveraging a triplet-loss function, DML exponentially increases training data compared to a traditional DL model. In this study, we investigated the utility of DML to improve the accuracy of DL models trained to classify cancerous lesions found on screening mammograms. Methods: Using a dataset of 2620 lesions found on routine screening mammogram, we trained both a traditional DL and DML models to classify suspicious lesions as cancerous or benign. The VGG16 architecture was used as the basis for the DL and DML models. Model performance was compared by calculating model accuracy, sensitivity, and specificity on a blinded test set of 378 lesions. In addition to individual model performance, we also measured agreement accuracy when both the DL and DML models were combined. Sub-analyses were conducted to identify phenotypes which were best suited for each model type. Both models underwent hyperparameters optimization to identify ideal batch size, learning rate, and regularization to prevent overfitting. Results: We found that the combination of the traditional DL model with DML model resulted in the highest overall accuracy (78.7%) representing a 7.1% improvement compared to the traditional DL model (p<.001). Alone, the traditional DL model had an improved accuracy compared to the DML model (71.4% vs 66.4%). The traditional DL model had a higher sensitivity (94.8% vs 73.6 %) , but lower specificity (34.7% vs 55.1%) compared the DML model. Sub-analyses suggested the traditional DL model was more accurate on higher density breasts, whereas the DML model was more accurate on lower density breasts. Additionally, the traditional DL model had the highest accuracy on oval shaped lesions, compared to the DML model which was most accurate on irregularly shaped breast lesions. Conclusion: Our study suggests that addition of DML models with traditional DL models can improve diagnostic image classification performance in cancer. Our results suggest DML models may provide increased specificity and help with classification of unique populations often misclassified by traditional DL models. Further studied investigating the utility of DML on other cancer imaging tasks are necessary to successfully build more robust DL models in cancer imaging. Citation Format: Justin Du, Sachin Umrao, Enoch Chang, Marina Joel, Aidan Gilson, Guneet Janda, Rachel Choi, Yongfeng Hui, Sanjay Aneja. The utility of deep metric learning for breast cancer identification on mammographic images [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2021; 2021 Apr 10-15 and May 17-21. Philadelphia (PA): AACR; Cancer Res 2021;81(13_Suppl):Abstract nr 184.

The need for intubation in methanol-poisoned patients, if not predicted in time, can lead to irreparable complications and even death. Artificial intelligence (AI) techniques like machine learning (ML) and deep learning (DL) greatly aid in accurately predicting intubation needs for methanol-poisoned patients. So, our study aims to assess Explainable Artificial Intelligence (XAI) for predicting intubation necessity in methanol-poisoned patients, comparing deep learning and machine learning models. This study analyzed a dataset of 897 patient records from Loghman Hakim Hospital in Tehran, Iran, encompassing cases of methanol poisoning, including those requiring intubation (202 cases) and those not requiring it (695 cases). Eight established ML (SVM, XGB, DT, RF) and DL (DNN, FNN, LSTM, CNN) models were used. Techniques such as tenfold cross-validation and hyperparameter tuning were applied to prevent overfitting. The study also focused on interpretability through SHAP and LIME methods. Model performance was evaluated based on accuracy, specificity, sensitivity, F1-score, and ROC curve metrics. Among DL models, LSTM showed superior performance in accuracy (94.0%), sensitivity (99.0%), specificity (94.0%), and F1-score (97.0%). CNN led in ROC with 78.0%. For ML models, RF excelled in accuracy (97.0%) and specificity (100%), followed by XGB with sensitivity (99.37%), F1-score (98.27%), and ROC (96.08%). Overall, RF and XGB outperformed other models, with accuracy (97.0%) and specificity (100%) for RF, and sensitivity (99.37%), F1-score (98.27%), and ROC (96.08%) for XGB. ML models surpassed DL models across all metrics, with accuracies from 93.0% to 97.0% for DL and 93.0% to 99.0% for ML. Sensitivities ranged from 98.0% to 99.37% for DL and 93.0% to 99.0% for ML. DL models achieved specificities from 78.0% to 94.0%, while ML models ranged from 93.0% to 100%. F1-scores for DL were between 93.0% and 97.0%, and for ML between 96.0% and 98.27%. DL models scored ROC between 68.0% and 78.0%, while ML models ranged from 84.0% to 96.08%. Key features for predicting intubation necessity include GCS at admission, ICU admission, age, longer folic acid therapy duration, elevated BUN and AST levels, VBG_HCO3 at initial record, and hemodialysis presence. This study as the showcases XAI's effectiveness in predicting intubation necessity in methanol-poisoned patients. ML models, particularly RF and XGB, outperform DL counterparts, underscoring their potential for clinical decision-making.

Study question Can heatmaps generated by occlusion explain the patterns learned by deep learning (DL) models classifying the embryo viability in ART? Summary answer Occlusion experiments generate heatmaps that reveal which regions in frames of time-lapse video (TLV) are more discriminative for classification and prediction by the DL models. What is known already DL has widely been explored in ART for embryo selection. Depending upon input (video or image), different DL models classifying embryo viability are developed. However, whether the prediction is based on actual input features or random guessing is unknown. The embryo selection in ART is subjective. If the intention is using DL models’ prediction to transfer, freeze or discard the embryo, explanations of how they interpret embryonic development features brings transparency and trust. In other areas, heatmaps are used for explaining DL predictions. The heatmaps can be a tool to understand patterns learned by DL models for embryo selection. Study design, size, duration We trained two separate DL models for predicting the presence of fetal heartbeat for the transferred embryos. We further used occlusion generated heatmaps to explain the predictions. For training, retrospective data was used. The input dataset consisted of 136 TLVs and corresponding patient data for 132 participants (128: single embryo transfers and 8: double embryo transfer) from both IVF and ICSI treatment. Each video was assessed by an embryologist. Participants/materials, setting, methods DL models (A as ResNet–18, B as VGG16) are trained for predicting the presence of fetal heartbeat on a single frame extracted from TLV after day three or later. Model A has a better recall (0.7) compared to B (0.5). Heatmaps explain the reason behind models’ recall rate by visually representing patterns learned by them. Using occlusion filter size 30*30 with stride 14 and size 50*50 with stride 25, we generate heatmaps for both models. Main results and the role of chance The heatmaps generated using occlusion can represent visually the patterns discovered by the DL models when predicting the presence of a fetal heartbeat. Using occlusion filter size 30*30 with stride 14, we verified that Model B has lower recall because the heatmaps show that the model finds redundant features present outside the embryo region in many input frames. It could be interpreted that either the model has not learned relevant patterns or is more robust to noise. This representation of DL models equips us in better decision-making, whether to consider or discard the prediction or rather train the model further, preprocess training data or change network architecture. The heatmaps revealed that for frames where significant patterns learned by the models are within the embryo region, more weight was given to specific features like the inner cell mass, trophectoderm and some parts within the zona pellucida. Moreover, the heat maps generated using occlusion are independent of the underlying model’s architecture as the same experiment settings were used for both models. For occlusion filter size 50*50 with stride 25, the expanse of input regions (in or outside the embryo) considered relevant could be visualized for both models A and B. Limitations, reasons for caution Heatmaps generated by occluding input regions give a visual representation of features in individual frames not directly on videos. Explaining DL models by heatmaps besides occlusion, other techniques (Grad-Cam) exist but were not evaluated. Furthermore, there is no quantitative measure for evaluating whether heatmaps are a good explanation or not. Wider implications of the findings: The heatmaps make the patterns discovered by DL models visually recognized and bring forth the prominent portions of embryo regions. This will again improve understanding and trust in DL models’ predictions. Visual representation of DL models using heatmaps enables interpreting a prediction, performing model analysis and determining scope for improvement. Trial registration number Not applicable

ObjectiveMicrovascular invasion (MVI) is of great significance for the individualized treatment of hepatocellular carcinoma (HCC) and preoperative noninvasive prediction of MVI is still an urgent clinical problem. To explore the effects of different regions of interest (ROI) and image input dimensions on the performance of deep learning (DL) models, and to select the best result to develop and validate a DL model for preoperative prediction of MVI.Materials and methodsA total of 206 patients with pathologically confirmed HCC from three hospitals were retrospectively enrolled and divided into training, internal validation and external test set. Based on hepatobiliary phase images (HBP) of gadoxetic acid-enhanced MRI, 2D DL, 3D DL and 2.5D deep multi-instance learning (MIL) models were established. The receiver operating characteristic curve (ROC) was used to evaluate the predictive efficacy of the above models. Based on the optimal performance model, the T1WI-FS and T2WI-FS images were preprocessed correspondingly, and a multimodal prediction model including three sequences was constructed. The ROC, and decision curve were used to visualize the predictive ability of the model.ResultsCompared with 2D DL and 3D DL models, the 2.5D DL model based on all axial images of ROI had the highest performance, with the AUC values of 0.802 (95% CI, 0.669–0.936) and 0.759 (95% CI, 0.643–0.875) in the validation and test sets. The AUCs of the multimodal MRI model were 0.954 (95% CI, 0.920–0.989) in the training set, 0.857 (95% CI, 0.736–0.978) in the validation set, and 0.788 (95% CI, 0.681–0.895) in the test set.ConclusionThe DL model that selects all axial slices of intratumor and peritumor as input shows robust capability in predicting MVI, which is expected to help clinical decision-making of individualized treatment for HCC.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12885-025-14971-7.

Failure to predict stroke promptly may lead to delayed treatment, causing severe consequences like permanent neurological damage or death. Early detection using deep learning (DL) and machine learning (ML) models can enhance patient outcomes and mitigate the long-term effects of strokes. The aim of this study is to compare these models, exploring their efficacy in predicting stroke. This study analyzed a dataset comprising 663 records from patients hospitalized at Hazrat Rasool Akram Hospital in Tehran, Iran, including 401 healthy individuals and 262 stroke patients. A total of eight established ML (SVM, XGB, KNN, RF) and DL (DNN, FNN, LSTM, CNN) models were utilized to predict stroke. Techniques such as 10-fold cross-validation and hyperparameter tuning were implemented to prevent overfitting. The study also focused on interpretability through Shapley Additive Explanations (SHAP). The evaluation of model’s performance was based on accuracy, specificity, sensitivity, F1-score, and ROC curve metrics. Among DL models, LSTM showed superior sensitivity at 96.15%, while FNN exhibited better specificity (96.0%), accuracy (96.0%), F1-score (95.0%), and ROC (98.0%) among DL models. For ML models, RF displayed higher sensitivity (99.9%), accuracy (99.0%), specificity (100%), F1-score (99.0%), and ROC (99.9%). Overall, RF outperformed all models, while DL models surpassed ML models in most metrics except for RF. DL models (CNN, LSTM, DNN, FNN) achieved sensitivities from 93.0 to 96.15%, specificities from 80.0 to 96.0%, accuracies from 92.0 to 96.0%, F1-scores from 87.34 to 95.0%, and ROC scores from 95.0 to 98.0%. In contrast, ML models (KNN, XGB, SVM) showed sensitivities between 29.0% and 94.0%, specificities between 89.47% and 96.0%, accuracies between 71.0% and 95.0%, F1-scores between 44.0% and 95.0%, and ROC scores between 64.0% and 95.0%. This study demonstrates the efficacy of DL and ML models in predicting stroke, with the RF models outperforming all others in key metrics. While DL models generally surpassed ML models, RF’s exceptional performance highlights the potential of combining these technologies for early stroke detection, significantly improving patient outcomes by preventing severe consequences like permanent neurological damage or death.

ObjectivesBased on the current clinical routine, we aimed to develop a novel deep learning model to distinguish coronavirus disease 2019 (COVID-19) pneumonia from other types of pneumonia and validate it with a real-world dataset (RWD).MethodsA total of 563 chest CT scans of 380 patients (227/380 were diagnosed with COVID-19 pneumonia) from 5 hospitals were collected to train our deep learning (DL) model. Lung regions were extracted by U-net, then transformed and fed to pre-trained ResNet-50-based IDANNet (Identification and Analysis of New covid-19 Net) to produce a diagnostic probability. Fivefold cross-validation was employed to validate the application of our model. Another 318 scans of 316 patients (243/316 were diagnosed with COVID-19 pneumonia) from 2 other hospitals were enrolled prospectively as the RWDs to testify our DL model’s performance and compared it with that from 3 experienced radiologists.ResultsA three-dimensional DL model was successfully established. The diagnostic threshold to differentiate COVID-19 and non-COVID-19 pneumonia was 0.685 with an AUC of 0.906 (95% CI: 0.886–0.913) in the internal validation group. In the RWD cohort, our model achieved an AUC of 0.868 (95% CI: 0.851–0.876) with the sensitivity of 0.811 and the specificity of 0.822, non-inferior to the performance of 3 experienced radiologists, suggesting promising clinical practical usage.ConclusionsThe established DL model was able to achieve accurate identification of COVID-19 pneumonia from other suspected ones in the real-world situation, which could become a reliable tool in clinical routine.Key Points• In an internal validation set, our DL model achieved the best performance to differentiate COVID-19 from non-COVID-19 pneumonia with a sensitivity of 0.836, a specificity of 0.800, and an AUC of 0.906 (95% CI: 0.886–0.913) when the threshold was set at 0.685.• In the prospective RWD cohort, our DL diagnostic model achieved a sensitivity of 0.811, a specificity of 0.822, and AUC of 0.868 (95% CI: 0.851–0.876), non-inferior to the performance of 3 experienced radiologists.• The attention heatmaps were fully generated by the model without additional manual annotation and the attention regions were highly aligned with the ROIs acquired by human radiologists for diagnosis.Supplementary InformationThe online version contains supplementary material available at 10.1007/s00330-020-07553-7.

PurposeTo develop generative adversarial networks (GANs) that synthesize realistic anterior segment optical coherence tomography (AS-OCT) images and evaluate deep learning (DL) models that are trained on real and synthetic datasets for detecting angle closure.MethodsThe GAN architecture was adopted and trained on the dataset with AS-OCT images collected from the Joint Shantou International Eye Center of Shantou University and the Chinese University of Hong Kong, synthesizing open- and closed-angle AS-OCT images. A visual Turing test with two glaucoma specialists was performed to assess the image quality of real and synthetic images. DL models, trained on either real or synthetic datasets, were developed. Using the clinicians’ grading of the AS-OCT images as the reference standard, we compared the diagnostic performance of open-angle vs. closed-angle detection of DL models and the AS-OCT parameter, defined as a trabecular-iris space area 750 µm anterior to the scleral spur (TISA750), in a small independent validation dataset.ResultsThe GAN training included 28,643 AS-OCT anterior chamber angle (ACA) images. The real and synthetic datasets for DL model training have an equal distribution of open- and closed-angle images (all with 10,000 images each). The independent validation dataset included 238 open-angle and 243 closed-angle AS-OCT ACA images. The image quality of real versus synthetic AS-OCT images was similar, as assessed by the two glaucoma specialists, except for the scleral spur visibility. For the independent validation dataset, both DL models achieved higher areas under the curve compared with TISA750. Two DL models had areas under the curve of 0.97 (95% confidence interval, 0.96–0.99) and 0.94 (95% confidence interval, 0.92–0.96).ConclusionsThe GAN synthetic AS-OCT images appeared to be of good quality, according to the glaucoma specialists. The DL models, trained on all-synthetic AS-OCT images, can achieve high diagnostic performance.Translational RelevanceThe GANs can generate realistic AS-OCT images, which can also be used to train DL models.

This paper aimed to develop and validate a deep learning (DL) model for automated detection of the laterality of the eye on anterior segment photographs. Anterior segment photographs for training a DL model were collected with the Scheimpflug anterior segment analyzer. We applied transfer learning and fine-tuning of pre-trained deep convolutional neural networks (InceptionV3, VGG16, MobileNetV2) to develop DL models for determining the eye laterality. Testing datasets, from Scheimpflug and slit-lamp digital camera photography, were employed to test the DL model, and the results were compared with a classification performed by human experts. The performance of the DL model was evaluated by accuracy, sensitivity, specificity, operating characteristic curves, and corresponding area under the curve values. A total of 14,468 photographs were collected for the development of DL models. After training for 100 epochs, the DL models of the InceptionV3 mode achieved the area under the receiver operating characteristic curve of 0.998 (with 95% CI 0.924–0.958) for detecting eye laterality. In the external testing dataset (76 primary gaze photographs taken by a digital camera), the DL model achieves an accuracy of 96.1% (95% CI 91.7%–100%), which is better than an accuracy of 72.3% (95% CI 62.2%–82.4%), 82.8% (95% CI 78.7%–86.9%) and 86.8% (95% CI 82.5%–91.1%) achieved by human graders. Our study demonstrated that this high-performing DL model can be used for automated labeling for the laterality of eyes. Our DL model is useful for managing a large volume of the anterior segment images with a slit-lamp camera in the clinical setting.

This is a single-period, single-product inventory model with several individual sources of demand. It is a multi-location problem with an opportunity for centralization. The holding and penalty cost functions at each location are assumed to be identical. Two types of inventory system are considered in this paper: the decentralized system and the centralized system. The decentralized system is a system in which a separate inventory is kept to satisfy the demand at each source of demand. The centralized system is a system in which all demands are satisfied from one central warehouse. This paper demonstrates that, for any probability distribution of a location's demands, the following properties are always true: given that the holding and penalty cost functions are identical at all locations, (1) if the holding and penalty cost functions are concave functions, then the expected holding and penalty costs in a decentralized system exceed those in a centralized system, except that (2) if the holding and penalty cost functions are linear functions, and for any i≠j, Pij, the coefficient of correlation between the ith location's demand and the jth location's demand is equal to 1, then the expected holding and penalty costs in a decentralized system are equal to those in a centralized system.

This is a single-period, single-product inventory model with several individual sources of demand. It is a multi-location problem with an opportunity for centralization. The holding and penalty cost functions at each location are assumed to be identical. Two types of inventory system are considered in this paper: the decentralized system and the centralized system. The decentralized system is a system in which a separate inventory is kept to satisfy the demand at each source of demand. The centralized system is a system in which all demands are satisfied from one central warehouse. This paper demonstrates that, for any probability distribution of a location's demands, the following properties are always true: given that the holding and penalty cost functions are identical at all locations, (1) if the holding and penalty cost functions are concave functions, then the expected holding and penalty costs in a decentralized system exceed those in a centralized system, except that (2) if the holding and penalty cost functions are linear functions, and for any i≠j, Pij, the coefficient of correlation between the ith location's demand and the jth location's demand is equal to 1, then the expected holding and penalty costs in a decentralized system are equal to those in a centralized system.

ABSTRACTSignificant advances have been made towards fault detection using deep learning. However, the fault labelling of seismic data requires great human effort. The resulting small sample problem makes traditional deep learning methods difficult to achieve desired results. Existing research proposes to train a deep learning model with labelled synthetic seismic data to get good fault detection results. However, due to the complexity of the actual geological situation, there are inevitable differences between synthetic seismic data and real seismic data in many aspects such as seismic signal frequency, frequency of fault distribution and degree of noise disturbance, which lead to the fact that the deep learning model trained by synthetic seismic data is difficult to get good fault detection result in field data applications. We propose to use transfer learning to reduce the impact of data differences to solve this problem: part of the deep transfer learning model is used to learn fault‐related features. And the other part of the deep transfer learning model is used to mine common features between the real seismic data and the synthetic seismic data, which makes the deep transfer learning model more suitable for real seismic data. Compared with the latest research progress, our method can greatly improve the effect of fault detection without real data label, which can significantly save the cost of manual label processing.

PurposeTo conduct a head-to-head comparison between deep learning (DL) and radiomics models across institutions for predicting microvascular invasion (MVI) in hepatocellular carcinoma (HCC) and to investigate the model robustness and generalizability through rigorous internal and external validation.MethodsThis retrospective study included 2304 preoperative images of 576 HCC lesions from two centers, with MVI status determined by postoperative histopathology. We developed DL and radiomics models for predicting the presence of MVI using B-mode ultrasound, contrast-enhanced ultrasound (CEUS) at the arterial, portal, and delayed phases, and a combined modality (B + CEUS). For radiomics, we constructed models with enlarged vs. original regions of interest (ROIs). A cross-validation approach was performed by training models on one center’s dataset and validating the other, and vice versa. This allowed assessment of the validity of different ultrasound modalities and the cross-center robustness of the models. The optimal model combined with alpha-fetoprotein (AFP) was also validated. The head-to-head comparison was based on the area under the receiver operating characteristic curve (AUC).ResultsThirteen DL models and 25 radiomics models using different ultrasound modalities were constructed and compared. B + CEUS was the optimal modality for both DL and radiomics models. The DL model achieved AUCs of 0.802–0.818 internally and 0.667–0.688 externally across the two centers, whereas radiomics achieved AUCs of 0.749–0.869 internally and 0.646–0.697 externally. The radiomics models showed overall improvement with enlarged ROIs (P < 0.05 for both CEUS and B + CEUS modalities). The DL models showed good cross-institutional robustness (P > 0.05 for all modalities, 1.6–2.1% differences in AUC for the optimal modality), whereas the radiomics models had relatively limited robustness across the two centers (12% drop-off in AUC for the optimal modality). Adding AFP improved the DL models (P < 0.05 externally) and well maintained the robustness, but did not benefit the radiomics model (P > 0.05).ConclusionCross-institutional validation indicated that DL demonstrated better robustness than radiomics for preoperative MVI prediction in patients with HCC, representing a promising solution to non-standardized ultrasound examination procedures.

ABSTRACTObjectivesPrevious deep learning studies have not explored the synergistic effects of ROI dimensions (2D/2.5D/3D), peritumoral expansion levels (0–8 mm), and segmentation scenarios (ROI only vs. ROI original). Our study aims to evaluate the performance of multidimensional deep transfer learning models in distinguishing molecular subtypes of breast cancer (luminal vs. nonluminal) using DCE‐MRI. Under two segmentation scenarios, we systematically compare the effects of ROI dimensions and peritumoral expansion levels to optimize multidimensional deep learning models via transfer learning for distinguishing luminal from nonluminal breast cancers in DCE‐MRI‐based analysis.Materials and MethodsFrom October 2020 to October 2023, data from 426 patients with primary invasive breast cancer were retrospectively collected. Patients were divided into three cohorts: (1) training cohort, n = 108, from SYSU Hospital (Zhuhai, China); (2) validation cohort 1, n = 165, from HZ Hospital (Huizhou, China); and (3) validation cohort 2, n = 153, from LY Hospital (Linyi, China). ROIs were delineated, and expansions of 2, 4, 6, and 8 mm beyond the lesion boundary were performed. We assessed the performance of various deep transfer learning models, considering precise segmentation (ROI only and ROI original) and varying peritumoral regions, using ROC curves and decision curve analysis.ResultsThe 2.5D1‐based deep learning model (ROI original, 4 mm expansion) demonstrated optimal performance, achieving an AUC of 0.808 (95% CI 0.715–0.901) in the training cohort, 0.766 (95% CI 0.682–0.850) in validation cohort 1, and 0.799 (95% CI 0.725–0.874) in validation cohort 2.ConclusionThe study highlights that the 2.5D1‐based deep learning model utilizing the three principal slices of the minimum bounding box (ROI original) with a 4 mm peritumoral region is effective in distinguishing between luminal and nonluminal breast cancer tumors, serving as a potential diagnostic tool.

To build and validate deep learning and machine learning fusion models to classify benign, malignant, and intermediate bone tumors based on patient clinical characteristics and conventional radiographs of the lesion. In this retrospective study, data were collected with pathologically confirmed diagnoses of bone tumors between 2012 and 2019. Deep learning and machine learning fusion models were built to classify tumors as benign, malignant, or intermediate using conventional radiographs of the lesion and potentially relevant clinical data. Five radiologists compared diagnostic performance with and without the model. Diagnostic performance was evaluated using the area under the curve (AUC). A total of 643 patients' (median age, 21 years; interquartile range, 12-38 years; 244 women) 982 radiographs were included. In the test set, the binary category classification task, the radiological model of classification for benign/not benign, malignant/nonmalignant, and intermediate/not intermediate had AUCs of 0.846, 0.827, and 0.820, respectively; the fusion models had an AUC of 0.898, 0.894, and 0.865, respectively. In the three-category classification task, the radiological model achieved a macro average AUC of 0.813, and the fusion model had a macro average AUC of 0.872. In the observation test, the mean macro average AUC of all radiologists was 0.819. With the three-category classification fusion model support, the macro AUC improved by 0.026. We built, validated, and tested deep learning and machine learning models that classified bone tumors at a level comparable with that of senior radiologists. Model assistance may somewhat help radiologists' differential diagnoses of bone tumors. • The deep learning model can be used to classify benign, malignant, and intermediate bone tumors. • The machine learning model fusing information from radiographs and clinical characteristics can improve the classification capacity for bone tumors. • The diagnostic performance of the fusion model is comparable with that of senior radiologists and is potentially useful as a complement to radiologists in a bone tumor differential diagnosis.