This repository contains a fine-tuned Densenet-201 model for breast cancer classification based on mammography images.

Due to the indistinguishable nature of the dataset various runs had been conducted to perform the original 3 classes classification according to the original DDSM dataset but the accuracy obtained is dismal (approx 67%) contrary to literature review of (>90%).

I have also explored dual input Swin Transformer using the Tumour Mask, however, similar dismal accuracy is obtained. We can look at the dataset and notice that the images all looks about the same except Normal. Thus, the detection strategy becomes detecting the presence of cancer by merging to Benign and Cancer images as a class against the Normal images.

With such approach, accuracy significant increases and achieve reliable performance.

The model is based on the Densenet-201 architecture, fine-tuned on the Mini-DDBS-JPEG dataset for breast cancer classification.

The model was trained with class balancing techniques to handle data imbalance. Performance metrics on the test set:

For detailed performance metrics including precision, recall, and F1-score per class, please check the training notebook.

Please check the inference compute.