We present a deep convolutional neural network for breast cancer screening exam classification, trained, and evaluated on over 200000 exams (over 1000000 images). Our network achieves an AUC of 0.895 in predicting the presence of cancer in the breast, when tested on the screening population. We attribute the high accuracy to a few technical advances. 1) Our network's novel two-stage architecture and training procedure, which allows us to use a high-capacity patch-level network to learn from pixel-level labels alongside a network learning from macroscopic breast-level labels. 2) A custom ResNet-based network used as a building block of our model, whose balance of depth and width is optimized for high-resolution medical images. 3) Pretraining the network on screening BI-RADS classification, a related task with more noisy labels. 4) Combining multiple input views in an optimal way among a number of possible choices. To validate our model, we conducted a reader study with 14 readers, each reading 720 screening mammogram exams, and show that our model is as accurate as experienced radiologists when presented with the same data. We also show that a hybrid model, averaging the probability of malignancy predicted by a radiologist with a prediction of our neural network, is more accurate than either of the two separately. To further understand our results, we conduct a thorough analysis of our network's performance on different subpopulations of the screening population, the model's design, training procedure, errors, and properties of its internal representations. Our best models are publicly available at https://github.com/nyukat/breast_cancer_classifier.

Purpose: The aim of this study was to investigate the utility of apparent diffusion coefficient (ADC) for prediction and early detection of treatment response in head and neck squamous cell carcinomas (HNSCC). Experimental Design: Diffusion-weighted magnetic resonance imaging studies were performed on 40 patients with newly diagnosed HNSCC before, during, and after the end of chemoradiation therapy. Analysis was done on data from 33 patients after exclusion of 7 patients that had incomplete data. Results: Pretreatment ADC value of complete responders (1.04 ± 0.19 × 10⁻³ mm²/s) was significantly lower (P < 0.05) than that from partial responders (1.35 ± 0.30 × 10⁻³ mm²/s). A significant increase in ADC was observed in complete responders within 1 week of treatment (P < 0.01), which remained high until the end of the treatment. The complete responders also showed significantly higher increase in ADC than the partial responders by the first week of chemoradiation (P < 0.01). When pretreatment ADC value was used for predicting treatment response, the area under the receiver operating characteristic curve was 0.80 with a sensitivity of 65% and a specificity of 86%. However, change in ADC within the first week of chemoradiation therapy resulted in an area under the receiver operating characteristic curve of 0.88 with 86% sensitivity and 83% specificity for prediction of treatment response. Conclusions: These results suggest that ADC can be used as a marker for prediction and early detection of response to concurrent chemoradiation therapy in HNSCC.

BACKGROUND AND PURPOSE: Tumor microenvironment, including blood flow and permeability, may provide crucial information regarding response to chemoradiation therapy. Thus, the objective of this study was to investigate the efficacy of pretreatment DCE-MR imaging for prediction of response to chemoradiation therapy in HNSCC. MATERIALS AND METHODS: DCE-MR imaging studies were performed on 33 patients with newly diagnosed HNSCC before neoadjuvant chemoradiation therapy by using a 1.5T (n = 24) or a 3T (n = 9) magnet. The data were analyzed by using SSM for estimation of K^trans, v_e, and τ_i. Response to treatment was determined on completion of chemoradiation as CR, with no evidence of disease (clinically or pathologically), or PR, with pathologically proved residual tumor. RESULTS: The average pretreatment K^trans value of the CR group (0.64 ± 0.11 minutes⁻¹, n = 24) was significantly higher (P = .001) than that of the PR (0.21 ± 0.05 minutes⁻¹, n = 9) group. No significant difference was found in other pharmacokinetic model parameters: v_e and τ_i, between the 2 groups. Although the PR group had larger metastatic nodal volume than the CR group, it was not significantly different (P = .276). CONCLUSIONS: These results indicate that pretreatment DCE-MR imaging can be potentially used for prediction of response to chemoradiation therapy of HNSCC.