Convolutional Neural Network Scores Research Articles

Effects of vitamin D supplementation on a deep learning-based mammographic evaluation in SWOG S0812.

Deep learning-based mammographic evaluations could noninvasively assess response to breast cancer (BC) chemoprevention. We evaluated change in a convolutional neural network (CNN)-based BC risk model applied to mammograms among women enrolled in SWOG S0812, which randomized 208 premenopausal high-risk women to receive oral vitamin D3 20,000IU weekly or placebo for 12 months. We applied the CNN model to mammograms collected at baseline (n = 109), 12 months (n = 97) and 24 months (n = 67), and compared changes in CNN risk score between treatment groups. Change in CNN score was neither significantly different between vitamin D and placebo groups at 12 months (0.005 vs. 0.002, p = 0.875) nor at 24 months (0.020 vs. 0.001, p = 0.563). The findings are consistent with the primary analysis of S0812, which did not demonstrate significant changes in MD with vitamin D supplementation compared to placebo. There is an ongoing need to evaluate biomarkers of response to novel BC chemopreventive agents.

When Spectral Modeling Meets Convolutional Networks: A Method for Discovering Reionization-era Lensed Quasars in Multiband Imaging Data

Over the last two decades, around 300 quasars have been discovered at z ≳ 6, yet only one has been identified as being strongly gravitationally lensed. We explore a new approach—enlarging the permitted spectral parameter space, while introducing a new spatial geometry veto criterion—which is implemented via image-based deep learning. We first apply this approach to a systematic search for reionization-era lensed quasars, using data from the Dark Energy Survey, the Visible and Infrared Survey Telescope for Astronomy Hemisphere Survey, and the Wide-field Infrared Survey Explorer. Our search method consists of two main parts: (i) the preselection of the candidates, based on their spectral energy distributions (SEDs), using catalog-level photometry; and (ii) relative probability calculations of the candidates being a lens or some contaminant, utilizing a convolutional neural network (CNN) classification. The training data sets are constructed by painting deflected point-source lights over actual galaxy images, to generate realistic galaxy–quasar lens models, optimized to find systems with small image separations, i.e., Einstein radii of θ E ≤ 1″. Visual inspection is then performed for sources with CNN scores of P lens > 0.1, which leads us to obtain 36 newly selected lens candidates, which are awaiting spectroscopic confirmation. These findings show that automated SED modeling and deep learning pipelines, supported by modest human input, are a promising route for detecting strong lenses from large catalogs, which can overcome the veto limitations of primarily dropout-based SED selection approaches.

Assessing electrocardiogram changes after ischemic stroke with artificial intelligence.

Ischemic stroke (IS) with subsequent cerebrocardiac syndrome (CCS) has a poor prognosis. We aimed to investigate electrocardiogram (ECG) changes after IS with artificial intelligence (AI). We collected ECGs from a healthy population and patients with IS, and then analyzed participant demographics and ECG parameters to identify abnormal features in post-IS ECGs. Next, we trained the convolutional neural network (CNN), random forest (RF) and support vector machine (SVM) models to automatically detect the changes in the ECGs; Additionally, We compared the CNN scores of good prognosis (mRS ≤ 2) and poor prognosis (mRS > 2) to assess the prognostic value of CNN model. Finally, we used gradient class activation map (Grad-CAM) to localize the key abnormalities. Among the 3506 ECGs of the IS patients, 2764 ECGs (78.84%) led to an abnormal diagnosis. Then we divided ECGs in the primary cohort into three groups, normal ECGs (N-Ns), abnormal ECGs after the first ischemic stroke (A-ISs), and normal ECGs after the first ischemic stroke (N-ISs). Basic demographic and ECG parameter analyses showed that heart rate, QT interval, and P-R interval were significantly different between 673 N-ISs and 3546 N-Ns (p < 0.05). The CNN has the best performance among the three models in distinguishing A-ISs and N-Ns (AUC: 0.88, 95%CI = 0.86-0.90). The prediction scores of the A-ISs and N-ISs obtained from the all three models are statistically different from the N-Ns (p < 0.001). Futhermore, the CNN scores of the two groups (mRS > 2 and mRS ≤ 2) were significantly different (p < 0.05). Finally, Grad-CAM revealed that the V4 lead may harbor the highest probability of abnormality. Our study showed that a high proportion of post-IS ECGs harbored abnormal changes. Our CNN model can systematically assess anomalies in and prognosticate post-IS ECGs.

The development of a convolutional neural network for the automatic detection of Northern Bobwhite Colinus virginianus covey calls

AbstractPassive acoustic monitoring using Autonomous Recording Units (ARUs) is becoming a significant research tool for collecting large amounts of ecological data. Northern bobwhite Colinus virginianus is an economically important game bird whose declining populations are of conservation concern, so efforts to monitor bobwhite abundance using ARUs are being intensified. Yet, manual processing of ARU data is time consuming and often expensive, so developing automatic call detection methods is a key step in acoustic monitoring. We present here the first single species convolutional neural network (CNN) developed purely for automatic bobwhite covey call identification and classification. We demonstrate the value of meaningful data augmentation by including non‐target calls and background noise into our training dataset, as well as evaluating alternative CNN score thresholds and model extrapolation performance. We trained our CNN on 6,682 manually labeled covey calls across three groups of sites within the southeastern USA. Precision and AUC from both CNN classification and individual call detection was high (0.80–0.99), and our model showed strong extrapolation ability across site groups. However, extrapolation performance significantly decreased for sites that were more dissimilar to the training data set if our meaningful data augmentation process was omitted. Our CNN detected significantly more covey calls than manual labeling using Raven Pro software, and processing time was greatly reduced: a single one hour wav file can be now analyzed by the CNN in roughly eight seconds. We also demonstrate using a simple case study that extremely high variability in estimates of bobwhite site occupancy and detection are obtained depending on the method of acoustic data processing (manual versus CNN). Our results suggest that our CNN provides robust and time‐saving analysis of bobwhite covey call acoustic data and can be applied to future research and monitoring projects with high confidence in the performance of the model.

Use of a convolutional neural network-based mammographic evaluation to predict breast cancer recurrence among women with hormone receptor-positive operable breast cancer.

We evaluated whether a novel, fully automated convolutional neural network (CNN)-based mammographic evaluation can predict breast cancer relapse among women with operable hormone receptor (HR)-positive breast cancer. We conducted a retrospective cohort study among women with stage I-III, HR-positive unilateral breast cancer diagnosed at Columbia University Medical Center from 2007 to 2017, who received adjuvant endocrine therapy and had at least two mammograms (baseline, annual follow-up) of the contralateral unaffected breast for CNN analysis. We extracted demographics, clinicopathologic characteristics, breast cancer treatments, and relapse status from the electronic health record. Our primary endpoint was change in CNN risk score (range, 0-1). We used two-samplet-tests to assess for difference in mean CNN scores between patients who relapsed vs. remained in remission, and conducted Cox regression analyses to assess for association between change in CNN score and breast cancer-free interval (BCFI), adjusting for known prognostic factors. Among 848 women followed for a median of 59months, there were 67 (7.9%) breast cancer relapses (36 distant, 25 local, 6 new primaries). There was a significant difference in mean absolute change in CNN risk score from baseline to 1-year follow-up between those who relapsed vs. remained in remission (0.001 vs. -0.022, p = 0.030). After adjustment for prognostic factors, a 0.01 absolute increase in CNN score at 1-year was significantly associated with BCFI, hazard ratio = 1.05 (95% Confidence Interval 1.01-1.09, p = 0.011). Short-term change in the CNN-based breast cancer risk model on adjuvant endocrine therapy predicts breast cancer relapse, and warrants further evaluation in prospective studies.

Unsupervised End-to-End Brain Tumor Magnetic Resonance Image Registration Using RBCNN: Rigid Transformation, B-Spline Transformation and Convolutional Neural Network.

Image registration is the process of aligning two or more images in a single coordinate. Nowadays, medical image registration plays a significant role in computer-assisted disease diagnosis, treatment, and surgery. The different modalities available in the medical image make medical image registration an essential step in Computer Assisted Diagnosis (CAD), Computer- Aided Therapy (CAT) and Computer-Assisted Surgery (CAS). Recently, many learning-based methods were employed for disease detection and classification, but those methods were not suitable for real-time due to delayed response and the need for pre-alignment and labeling. The proposed research constructed a deep learning model with Rigid transform and B-Spline transform for medical image registration for an automatic brain tumour finding. The proposed research consists of two steps. The first step uses Rigid transformation based Convolutional Neural Network and the second step uses B-Spline transform-based Convolutional Neural Network. The model is trained and tested with 3624 MR (Magnetic Resonance) images to assess the performance. The researchers believe that MR images help in the success of the treatment of patients with brain tumour. The result of the proposed method is compared with the Rigid Convolutional Neural Network (CNN), Rigid CNN + Thin-Plat Spline (TPS), Affine CNN, Voxel morph, ADMIR (Affine and Deformable Medical Image Registration) and ANT(Advanced Normalization Tools) using DICE score, Average Symmetric surface Distance (ASD), and Hausdorff distance. The RBCNN model will help the physician to automatically detect and classify the brain tumor quickly (18 Sec) and efficiently without doing pre-alignment and labeling.

Use of a novel convolutional neural network-based mammographic evaluation to assess response to adjuvant endocrine therapy in women with early-stage breast cancer.

530 Background: The standard of care for early-stage hormone receptor (HR)-positive breast cancer (BC) is 5-10 years of adjuvant endocrine therapy (ET), which leads to a 50-60% relative risk reduction in BC recurrence. However, 10-40% of patients may relapse up to 20 years (y) after diagnosis, and there is a need for biomarkers of response to ET. We developed a novel, fully-automated convolutional neural network (CNN)-based mammographic evaluation that accurately predicts BC risk, which is being evaluated as a pharmacodynamic response biomarker to adjuvant ET. Methods: We conducted a retrospective cohort study among women with HR-positive stage I-III unilateral BC diagnosed at Columbia University Irving Medical Center from 2007-2017, who received adjuvant ET and had at least 2 mammograms of the contralateral breast (baseline and annual follow-up). Demographics, clinical characteristics, BC treatments, and relapse status were extracted from the electronic health record and New York-Presbyterian Hospital Tumor Registry. We performed CNN analysis of mammograms at baseline (start of ET) and annual follow-up. Our primary endpoint was change in CNN risk score, expressed as a continuous variable (range, 0-1). We used two-sample t-tests to assess for differences in mean CNN scores between patients who relapsed or remained in remission. We evaluated if CNN score at baseline and change from baseline were associated with relapse using logistic regression, with adjustment for known prognostic factors. Results: Among 870 evaluable women, mean age at diagnosis was 59.5y (standard deviation [SD], 12.4); 60.3% had stage I tumors, 72.6% underwent lumpectomy, and 45.8% received chemotherapy. With a median follow-up of 4.9y, there were 68 (7.9%) breast cancer relapses (36 distant, 26 local, 6 new primary). Median number of evaluable mammograms per patient was 5 (range, 2-13). Mean baseline CNN risk scores were significantly higher among women who relapsed compared to those in remission (0.258 vs 0.237, p = 0.022), which remained significant after adjustment for known prognostic factors. There was a significant difference in mean absolute change in CNN risk score from baseline to 1y follow-up between those who relapsed vs. remained in remission (0.001 vs. -0.022, p = 0.027), but this was no longer significant in multivariable analysis. Conclusions: We demonstrated that higher baseline CNN risk score was an independent predictor of BC relapse. A greater decrease in mean CNN risk scores at 1-year follow-up after initiating adjuvant ET was seen among BC patients who remained in remission compared to those who relapsed. Therefore, baseline CNN risk scores may identify patients at high-risk for breast cancer recurrence to target for more intensive adjuvant treatment. Early changes in CNN risk scores may be used to predict response to long-term ET in the adjuvant setting.

Abstract PS5-21: Use of a novel convolutional neural network (CNN)-based mammographic evaluation to assess response to adjuvant endocrine therapy

Abstract Background: The standard of care for early-stage hormone receptor (HR)-positive breast cancer is 5-10 years of adjuvant endocrine therapy (ET), which is associated with a 50-60% relative risk reduction in breast cancer recurrence. However, patients remain at risk of recurrence up to 20 years after diagnosis, and there is a need for biomarkers of response to ET. While decrease in mammographic density (MD) is associated with improved disease-free survival (DFS), measurement is subject to variability in radiologists’ interpretations. We developed a novel, fully-automated convolutional neural network (CNN)-based mammographic evaluation that is a more accurate, independent predictor of breast cancer risk than MD. We evaluated the role of the CNN model as a pharmacodynamic biomarker of response to adjuvant ET among women with early-stage HR-positive breast cancer. Methods: We conducted a retrospective cohort study among women with HR-positive, stage I-III unilateral breast cancer diagnosed at Columbia University Irving Medical Center (CUIMC) in New York, NY, from 2007-2017, who received adjuvant ET and had at least two contralateral mammograms (baseline and on ET) in our electronic health record (EHR). Demographics, clinical characteristics, breast cancer treatment (surgery, radiation, systemic therapy), type of ET (aromatase inhibitor [AI], tamoxifen, or both), and breast cancer relapse (distant, local, new breast primary) were extracted from the EHR and New York Presbyterian Hospital (NYPH) Tumor Registry. We performed CNN analysis of contralateral mammograms at baseline (within 1 year of diagnosis and prior to ET) and at 1- and 2-year (y) follow-up on ET. The primary endpoint was change in CNN risk score, expressed as a continuous variable (range, 0-1.00). Paired t-tests were used to assess for differences in CNN scores between baseline and 1y and 2y follow-up. Logistic regression was used to evaluate if CNN scores at baseline and change from baseline were associated with relapse, with adjustment for known prognostic factors. Results: Of 2,559 women diagnosed with stage I-III HR-positive breast cancer at CUIMC from 2007-2017, 465 had serial mammograms available for CNN analysis. Mean age at diagnosis was 61.2y (SD, 12.2y), and 38.1% of women were non-Hispanic white, 12.5% non-Hispanic black, 39.8% Hispanic, and 9.7% other/unknown. At initial diagnosis, 62.2% had stage I tumors, 74.4% received lumpectomy, and 41.7% received chemotherapy. There were 28 (6.0%) breast cancer relapses (15 distant, 10 local, 3 new primary). Women who had relapsed were more likely to be obese (p=0.009), have higher tumor stage (p&amp;lt;0.001) and grade (p=0.022), higher Ki67% (p=0.003), and have received chemotherapy (p=0.036). While baseline CNN risk scores were higher among women who had relapsed compared to those who did not (0.295 vs 0.269), this difference was not statistically significant (p=0.197). Among all women, mean CNN scores were 0.271 (SD, 0.101) at baseline, 0.216 (SD, 0.073) at 1y, and 0.208 (SD, 0.072) at 2y; changes in scores from baseline to 1y and 2y were significant (p&amp;lt;0.001), consistent across subgroups. However, in multivariable logistic regression, there was no significant association between change in CNN score from baseline to 1y and 2y and breast cancer relapse. Conclusions: We demonstrated a significant change in CNN risk scores from baseline to short-term follow-up among women with early-stage breast cancer who received adjuvant ET. Women who relapsed had higher baseline CNN risk scores, but this was not significant. Due to the small number of relapses, change in CNN risk scores was not associated with breast cancer recurrence. The CNN risk model will be evaluated prospectively in adjuvant clinical trials, to further evaluate its role as a predictor of breast cancer relapse and response to adjuvant ET. Citation Format: Julia E McGuinness, Jana Lee, Aishwarya Anuraj, Simukayi Mutasa, Richard S Ha, MD, Katherine D Crew, MDMS. Use of a novel convolutional neural network (CNN)-based mammographic evaluation to assess response to adjuvant endocrine therapy [abstract]. In: Proceedings of the 2020 San Antonio Breast Cancer Virtual Symposium; 2020 Dec 8-11; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2021;81(4 Suppl):Abstract nr PS5-21.

Convolutional neural network for discriminating nasopharyngeal carcinoma and benign hyperplasia on MRI.

A convolutional neural network (CNN) was adapted to automatically detect early-stage nasopharyngeal carcinoma (NPC) and discriminate it from benign hyperplasia on a non-contrast-enhanced MRI sequence for potential use in NPC screening programs. We retrospectively analyzed 412 patients who underwent T2-weighted MRI, 203 of whom had biopsy-proven primary NPC confined to the nasopharynx (stage T1) and 209 had benign hyperplasia without NPC. Thirteen patients were sampled randomly to monitor the training process. We applied the Residual Attention Network architecture, adapted for three-dimensional MR images, and incorporated a slice-attention mechanism, to produce a CNN score of 0-1 for NPC probability. Threefold cross-validation was performed in 399 patients. CNN scores between the NPC and benign hyperplasia groups were compared using Student's t test. Receiver operating characteristic with the area under the curve (AUC) was performed to identify the optimal CNN score threshold. In each fold, significant differences were observed in the CNN scores between the NPC and benign hyperplasia groups (p < .01). The AUCs ranged from 0.95 to 0.97 with no significant differences between the folds (p = .35 to .92). The combined AUC from all three folds (n = 399) was 0.96, with an optimal CNN score threshold of > 0.71, producing a sensitivity, specificity, and accuracy of 92.4%, 90.6%, and 91.5%, respectively, for NPC detection. Our CNN method applied to T2-weighted MRI could discriminate between malignant and benign tissues in the nasopharynx, suggesting that it as a promising approach for the automated detection of early-stage NPC. • The convolutional neural network (CNN)-based algorithm could automatically discriminate between malignant and benign diseases using T2-weighted fat-suppressed MR images. • The CNN-based algorithm had an accuracy of 91.5% with an area under the receiver operator characteristic curve of 0.96 for discriminating early-stage T1 nasopharyngeal carcinoma from benign hyperplasia. • The CNN-based algorithm had a sensitivity of 92.4% and specificity of 90.6% for detecting early-stage nasopharyngeal carcinoma.

MNet2FPGA: A Design Flow for Mapping a Fixed-Point CNN to Zynq SoC FPGA

The convolutional neural networks (CNNs) are a computation and memory demanding class of deep neural networks. The field-programmable gate arrays (FPGAs) are often used to accelerate the networks deployed in embedded platforms due to the high computational complexity of CNNs. In most cases, the CNNs are trained with existing deep learning frameworks and then mapped to FPGAs with specialized toolflows. In this paper, we propose a CNN core architecture called mNet2FPGA that places a trained CNN on a SoC FPGA. The processing system (PS) is responsible for convolution and fully connected core configuration according to the list of prescheduled instructions. The programmable logic holds cores of convolution and fully connected layers. The hardware architecture is based on the advanced extensible interface (AXI) stream processing with simultaneous bidirectional transfers between RAM and the CNN core. The core was tested on a cost-optimized Z-7020 FPGA with 16-bit fixed-point VGG networks. The kernel binarization and merging with the batch normalization layer were applied to reduce the number of DSPs in the multi-channel convolutional core. The convolutional core processes eight input feature maps at once and generates eight output channels of the same size and composition at 50 MHz. The core of the fully connected (FC) layer works at 100 MHz with up to 4096 neurons per layer. In a current version of the CNN core, the size of the convolutional kernel is fixed to 3×3. The estimated average performance is 8.6 GOPS for VGG13 and near 8.4 GOPS for VGG16/19 networks.

Music auto-tagging using scattering transform and convolutional neural network with self-attention

As a branch of machine learning, deep learning has been used for tackling with the music auto-tagging problem. Deep learning methods, especially those with convolutional neural network (CNN) architecture, have exhibited good performance on this multi-label classification task. However, the feature extracting part and preprocessing part of this architecture need to be improved. In this paper, we propose a deep-learning model based on CNN with scattering transform and self-attention mechanism for music automatic tagging. To get a balance between information integrity and feature extraction in the preprocessing phase, we employ the scattering transform. Then, a multi-layer CNN is used to extract higher-level features from the scattering coefficients. In order to select better receptive fields of the CNN, self-attention sub-network is appended at the last layer of CNN. Experimental results on the MagnaTagATune dataset and Million Song Dataset (MSD) show the proposed model is a good choice for music auto-tagging task, since the scores of the area under the receiver operating characteristic curve (ROC-AUC) and the area under the precision–recall curve (PR-AUC) obtained in this paper surpass the state-of-the-art models. Furthermore, we visualize the distributions of attention weights, activations of the CNN and ROC-AUC scores on each tag for better understanding of the model.

Volumetric assessment of paranasal sinus opacification on computed tomography can be automated using a convolutional neural network.

Computed tomography (CT) plays a key role in evaluation of paranasal sinus inflammation, but improved, and standardized, objective assessment is needed. Computerized volumetric analysis has benefits over visual scoring, but typically relies on manual image segmentation, which is difficult and time-consuming, limiting practical applicability. We hypothesized that a convolutional neural network (CNN) algorithm could perform automatic, volumetric segmentation of the paranasal sinuses on CT, enabling efficient, objective measurement of sinus opacification. In this study we performed initial clinical testing of a CNN for fully automatic quantitation of paranasal sinus opacification in the diagnostic workup of patients with chronic upper and lower airway disease. Sinus CT scans were collected on 690 patients who underwent imaging as part of multidisciplinary clinical workup at a tertiary care respiratory hospital between April 2016 and November 2017. A CNN was trained to perform automatic segmentation using a subset of CTs (n = 180) that were segmented manually. A nonoverlapping set (n = 510) was used for testing. CNN opacification scores were compared with Lund-MacKay (LM) visual scores, pulmonary function test results, and other clinical variables using Spearman correlation and linear regression. CNN scores were correlated with LM scores (rho = 0.82, p < 0.001) and with forced expiratory volume in 1 second (FEV1 ) percent predicted (rho = -0.21, p < 0.001), FEV1 /forced vital capacity ratio (rho = -0.27, p < 0.001), immunoglobulin E (rho = 0.20, p < 0.001), eosinophil count (rho = 0.28, p < 0.001), and exhaled nitric oxide (rho = 0.40, p < 0.001). Segmentation of the paranasal sinuses on CT can be automated using a CNN, providing truly objective, volumetric quantitation of sinonasal inflammation.

Abstract P1-02-06: Effect of chemoprevention on convolutional neural network-based breast cancer risk model using a mammographic dataset of women with atypical hyperplasia, lobular and ductal carcinoma in situ

Abstract Background: Mammographic density (MD) is a strong predictor of breast cancer risk and has been shown to predict response to tamoxifen for chemoprevention. We have developed a novel convolutional neural network (CNN)-derived pixel-wise breast cancer risk model, which is a more accurate predictor of breast cancer risk than MD. Our objective was to investigate how chemoprevention use among women with atypical hyperplasia (AH), lobular or ductal carcinoma in situ (LCIS/DCIS) is associated with change in the CNN breast cancer risk model. Methods: We conducted a retrospective cohort study of patients diagnosed with AH, LCIS, or DCIS at Columbia University Irving Medical Center (CUIMC) in New York, NY between 2007 and 2015. We collected mammograms conducted at baseline (at diagnosis of AH/LCIS/DCIS or start of chemoprevention) and 3-5 years later. Demographic and clinical characteristics, chemoprevention use, and mammography images were extracted from the electronic health record. Briefly, each mammogram was normalized as a map of z-scores and resized to an input image size of 256x256. Then a contracting and expanding fully convolutional CNN architecture was composed entirely of 3x3 convolutions, a total of four strided convolutions instead of pooling layers, and symmetric residual connections. L2 regularization and augmentation methods were implemented to prevent over-fitting. High-risk was defined as an absolute CNN model score &amp;gt;0.5. Change in CNN breast cancer risk model from baseline to follow-up was defined as: 1) stay high risk; 2) stay low risk; 3) increase from low to high risk; 4) decrease from high to low risk. Unordered polytomous regression models were used to assess the association between change in CNN score with chemoprevention use. Results: Among 541 evaluable patients, the mean age was 60 years (SD, 9.56) and 69% were postmenopausal. About a third of women (N=184, 34%) initiated chemoprevention, including 72% on a selective estrogen receptor modulator (SERM), 19% on an aromatase inhibitor (AI), and 9% both. About 57% had high CNN breast cancer risk scores at baseline. Comparing women who initiated chemoprevention to those who did not, more women had a decrease in their CNN breast cancer risk model (33.7% vs. 22.9%) and fewer women had an increase in CNN score (11.4% vs. 20.2%). In multivariate analysis, those who were treated with chemoprevention were less likely to have an increase in CNN risk score (odds ratio=0.51, 95% confidence interval=0.28-0.92). Conclusions: We previously reported that our CNN based algorithm can predict breast cancer risk and we now demonstrate that the CNN breast cancer risk model is modifiable on serial mammograms from high-risk women with proven chemopreventive agents, such as SERMs and AIs. Future studies should focus on whether these CNN mammographic changes are associated with subsequent development of breast cancer and may serve as a pharmacodynamics surrogate endpoint for breast cancer chemoprevention trials. Citation Format: Haley Manley, Simukayi Mutasa, Richard Ha, Katherine Crew. Effect of chemoprevention on convolutional neural network-based breast cancer risk model using a mammographic dataset of women with atypical hyperplasia, lobular and ductal carcinoma in situ [abstract]. In: Proceedings of the 2019 San Antonio Breast Cancer Symposium; 2019 Dec 10-14; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2020;80(4 Suppl):Abstract nr P1-02-06.

Half a Percent of Labels is Enough: Efficient Animal Detection in UAV Imagery Using Deep CNNs and Active Learning

We present an Active Learning (AL) strategy for re-using a deep Convolutional Neural Network (CNN)-based object detector on a new dataset. This is of particular interest for wildlife conservation: given a set of images acquired with an Unmanned Aerial Vehicle (UAV) and manually labeled gound truth, our goal is to train an animal detector that can be re-used for repeated acquisitions, e.g. in follow-up years. Domain shifts between datasets typically prevent such a direct model application. We thus propose to bridge this gap using AL and introduce a new criterion called Transfer Sampling (TS). TS uses Optimal Transport to find corresponding regions between the source and the target datasets in the space of CNN activations. The CNN scores in the source dataset are used to rank the samples according to their likelihood of being animals, and this ranking is transferred to the target dataset. Unlike conventional AL criteria that exploit model uncertainty, TS focuses on very confident samples, thus allowing a quick retrieval of true positives in the target dataset, where positives are typically extremely rare and difficult to find by visual inspection. We extend TS with a new window cropping strategy that further accelerates sample retrieval. Our experiments show that with both strategies combined, less than half a percent of oracle-provided labels are enough to find almost 80% of the animals in challenging sets of UAV images, beating all baselines by a margin.