Matthews Correlation Coefficient Loss for Deep Convolutional Networks: Application to Skin Lesion Segmentation

MCC Loss is now available in smp Segmentation Models PyTorch as smp.losses.MCCLoss.

This is the code corresponding to our ISBI 2021 paper. If you use our code, please cite our paper:

Kumar Abhishek, Ghassan Hamarneh, "Matthews Correlation Coefficient Loss for Deep Convolutional Networks: Application to Skin Lesion Segmentation", The IEEE International Symposium on Biomedical Imaging (ISBI), 2021.

The corresponding BibTeX entry is:

@InProceedings{Abhishek_2021_ISBI,
author = {Abhishek, Kumar and Hamarneh, Ghassan},
title = {Matthews Correlation Coefficient Loss for Deep Convolutional Networks: Application to Skin Lesion Segmentation},
booktitle = {The IEEE International Symposium on Biomedical Imaging (ISBI)},,
pages={225--229},
month = {April},
year = {2021}
}

Dependencies

• PyTorch

Usage

from loss import MCCLoss

# Default: probabilities in, batch-wise reduction
criterion = MCCLoss()
y_pred = torch.rand(4, 1, 128, 128)          # predicted probabilities
y_true = torch.randint(0, 2, (4, 1, 128, 128)).float()
loss = criterion(y_pred, y_true)

# With raw logits (before sigmoid)
criterion = MCCLoss(from_logits=True)
loss = criterion(logits, masks)

# Per-sample reduction (compute MCC per sample, then average)
criterion = MCCLoss(reduction="sample")

A simple example with 5×5 binary masks is shown in Example.ipynb.

Tests

python -m pytest test_mcc.py -v