Tiny Faces for Small Lesion Detection

This repo includes Peiyun Hu's awesome tiny face detector for the use case of small lesion detection in mammographic scans.

Getting Started

• Clone this repository.
• Download the DDSM and AIIMS dataset and annotations files to data/ folder.
• Install dependencies with conda env create -f env.yml.
• Run Preprocess_Data_Tiny_Faces.py to preprocess XMLs and images for WiderFace format.

Training

Just type make at the repo root and you should be good to go!

In case you wish to change some settings (such as data location), you can modify the Makefile which should be super easy to work with.

Evaluation

To run evaluation and generate the output files as per the WIDERFace specification, simply run make evaluate. The results will be stored in the val_results directory.

You can then use the dataset's eval_tools to generate the mAP numbers (this needs Matlab/Octave).

Similarly, to run the model on the test set, run make test to generate results in the test_results directory.

FROC Curve

Lesion classification and localization in medical imaging is usually done using FROC curve. To generate the FROC, use the FROC_Sumanyu_updated.py script.