Deep learning for breast cancer diagnosis in contrast-enhanced breast CT

Background: Breast cancer is the second most common cause of death from cancer in women in the United States after lung cancer. Thanks to early detection and treatment improvements, the mortality rate has been steadily decreasing in the last decades. Therefore, there is an increasing interest in finding new methodologies for improving the current state of the art. Several works validated the efficiency of Artificial Intelligence (AI) algorithms for cancer detection and diagnosis, but the application of uncertainty-based models, which have potential to enhance result interpretability and therefore clinical translation, remains to be investigated in depth. Project: This thesis aims to develop and validate Deep Learning algorithms for tumor classification and segmentation on 3D contrast-enhanced breast computed tomographic (CE-BCT) scans, exploring the mass-level uncertainty of the predictions through the Monte Carlo Dropout. Methods: 542 biopsy-proven breast masses (181 benign, 343 malignant) from 409 patients were imaged with a clinical Breast CT system after iodinated contrast medium administration. A 3D volume of interest (VOI) of 3.5cm per side was placed around each mass, and all masses were manually annotated in 3D by a board-certified breast radiologist. The mass VOIs and respective binary annotations were used to train (n = 262) and fine-tune (n = 88) a two-channel 3D Dense Convolutional Network and a 3D Residual UNet for mass classification and segmentation, respectively. Both networks were tested on an independent dataset of 192 biopsy-proven breast masses (89 benign, 103 malignant). The classification algorithm was evaluated with the area under the receiver operating characteristics curve (AUC), with 95% confidence interval (C.I.) calculated with bootstrapping (2,000 bootstraps) whereas the segmentation architecture was evaluated with the Dice score. Finally, multiple mass-level uncertainty metrics were tested on both classification and segmentation Monte Carlo outcomes, analyzing the performance improvement obtained by rejecting the predictions at different uncertainty and sensitivity thresholds. Results: On the independent test set, the two-channels 3D Dense Convolutional Network achieved an AUC of 0.84 (95% CI 0.78-0.90). Then, the 3D Residual UNet achieved an average DICE score of 0.79 ± 0.2. Finally, low-performance classification was found to be correlated with high variance, increasing the accuracy by 12% when 57 test masses with the highest prediction uncertainty were excluded. Moreover, the uncertainty was also found to be correlated with the segmentation performances, observing linear correlation coefficients (ρ) of 0.76 and 0.58 for the Intersection over Unions (IoUs) and the average Dice score over Monte Carlo samples, respectively. This allowed to increase the Dice score by 12% in both cases by removing 57 test masses based on their relative uncertainty metric. Conclusions: The AI methods developed and validated for this study achieved satisfactory performances and the evaluation of the uncertainty for the exclusion of the masses might enhance the performances. This could possibly be valuable for facilitating the translation of AI into clinics.