Classification of Precancerous Colorectal Lesions via ConvNeXt on Histopathological Images

Abstract

In this translational study, the classification of precancerous colorectal lesions is performed by the ConvNeXt method on MHIST histopathological imaging dataset. The ConvNeXt method is the modernized ResNet-50 architecture having some training tricks inspired by Swin Transformers and ResNeXT. The performance of the ConvNeXt models are benchmarked on different scenarios such as ‘full data’, ‘gradually increasing difficulty based data’ and ‘k-shot data’. The ConvNeXt models outperformed almost all the other studies which are applied on MHIST by using ResNet models, vision transformers, weight distillation, self-supervised learning and curriculum learning strategy in terms of different scenarios and metrics. The ConvNeXt model trained with ‘full data’ yields the best result with the score of 0.8890 for accuracy, 0.9391 for AUC, 0.9121 for F1 and 0.7633 for cohen’s cappa.