Deep learning for breast cancer diagnosis

Breast cancer is a common malignancy and a leading cause of cancer-related deaths in women worldwide. Its early diagnosis can significantly reduce morbidity and mortality rates in women. To this end, histopathological diagnosis is usually followed as a gold standard approach. However, this process is tedious, labor-intensive, and may be subject to inter-reader variability. Accordingly, an automatic diagnostic system can assist in improving the quality of diagnosis. The key intent of this thesis is to design, optimize, and validate end-to-end systems based on deep learning for the effective and efficient diagnosis of breast malignancy using histopathology images. In our first contribution, we conducted a systematic review of state-of-the-art supervised machine and deep learning approaches in the detection, segmentation, and classification of breast lesions using widely used medical imaging modalities including mammography, sonography, magnetic resonance imaging, and histopathology during the years 2016 and 2022. It is inferred that convolutional neural networks are progressively being exploited in computer-aided diagnosis of breast cancer. Furthermore, it is deduced that mammography and magnetic resonance imaging were mostly utilized in detection and segmentation tasks, followed by sonography. Whereas, mammography and histopathology were predominantly used in classification tasks. In our second contribution, we accomplished the first case study related to the binary classification of breast cancer. In this study, we presented an ensemble deep learning approach for the definite classification of non-carcinoma and carcinoma histopathology images using our collected dataset. We trained four different models based on pre-trained VGG16 and VGG19 architectures. Initially, we followed 5-fold crossvalidation operations on all the individual models, namely, fully-trained VGG16, finetuned VGG16, fully-trained VGG19, and fine-tuned VGG19 models. Then, we followed an ensemble strategy by taking the average of predicted probabilities and found that the ensemble of fine-tuned VGG16 and fine-tuned VGG19 performed competitive classification performance, especially on the carcinoma class. The ensemble of finetuned VGG16 and VGG19 models offered sensitivity of 97.73% for the carcinoma class. Moreover, it offered accuracy of 95.29% and F1-score of 95.29%. These experimental results demonstrated that our proposed deep learning approach is effective for the automatic classification of complex-natured histopathology images of breast cancer, more specifically for carcinoma images. In our third and final contribution, we accomplished the second case study related to the multiclass classification of breast cancer. In this study, we presented a deep learning approach to automatically classify hematoxylin-eosin-stained microscopy images into normal tissues, benign lesions, in situ carcinoma, and invasive carcinoma using our collected dataset. The proposed model exploited six intermediate layers of the Xception network to retrieve robust and abstract features from input images. First, we optimized the proposed model on the original (unnormalized) dataset using 5-fold cross-validation. Then, we investigated its performance on four normalized datasets resulting from Reinhard, Ruifrok, Macenko, and Vahadane stain normalization. For original images, the proposed framework yielded accuracy of 98.00% along with Cohen’s kappa score of 0.969. Furthermore, it achieved an average AUC-ROC score of 0.998 as well as a mean AUC-PR value of 0.995. Specifically, for in situ carcinoma and invasive carcinoma, it offered sensitivity of 96.00% and 99.00%, respectively. For normalized images, the proposed architecture performed better for Macenko nor malization compared to the other three techniques. In this case, the proposed model achieved accuracy of 97.79% together with Cohen’s kappa score of 0.965. In addition, it attained an average AUC-ROC score of 0.997 and a mean AUC-PR value of 0.991. Especially, for in situ carcinoma and invasive carcinoma, it offered sensitivity of 96.00% and 99.00%, respectively. These results demonstrate that our proposed model outperformed the baseline AlexNet as well as state-of-the-art VGG16, VGG19, Inception-v3, and Xception models with their default settings. Furthermore, it can be inferred that although stain normalization techniques offered competitive performance, they could not surpass the results of the original dataset. This thesis is accomplished with three journal articles whereas the fourth one is under review. Similarly, two papers are presented and published at the international conferences. Moreover, an international research stay was successfully performed at the Université Laval in Canada.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.