Breast Cancer Histopathological Image Classification using EfficientNet Architecture

Breast cancer is the most common type of cancer affecting women. The formation of lumps in the breast is one of the first signs of the presence of this disease. These tumors can either be cancerous or benign and hence a breast tissue biopsy is conducted to determine their nature. Advancements in the field of vision-based Deep Learning have facilitated the wide adoption of automated diagnostic systems in hospitals, for tasks such as cancer and COVID detection from lung X-ray scans, diabetic retinopathy detection from retinal fundus images, brain MRI segmentation, etc. Moving forward, reduction in training, validation and development times, and efficient usage of training resources for these models will be more in focus. The EfficientNet architecture proposed by Google has recently outperformed prior state-of-the-art architectures such as DenseNet and ResNet on the ImageNet classification task while using fewer parameters and epochs to converge faster. In this paper, we compare the performance of the EfficientNetB3 architecture with the above-mentioned architectures for the tasks of binary and multinomial tumor classification on the benchmark BreakHis dataset, which consists of around 8000 breast histopathology images of varying magnification. Our results show that under similar training conditions, the EfficientNetB3 can converge faster and outperform the previous benchmark models by a significant margin. Our best models achieved 100% sensitivity and accuracy on certain binary classification tasks and a sensitivity of 95.45% and precision of 95.15% on 8-ary classification tasks.