LIVER CANCER DETECTION USING RESNET, SWIN TRANSFORMER AND CONVNEXT TINY NETWORKS

Abstract

Liver tumor detection from CT images is challenging because most CT slices do not contain tumors, while only a small number contain tumor regions. Another challenge is that slices from the same patient can be very similar, which may cause data leakage when splitting the data randomly. This study focuses on classifying liver CT slices into Tumor and No-Tumor classes using deep learning. The LiTS and 3Dircadb datasets are combined and used for the experiments. Three deep learning models, Swin-T, ConvNeXt-Tiny, and ResNet-50, are trained and evaluated. A patient-level data split is used to keep patients in the training, validation, and test sets separate. The three models are also combined using a soft-voting ensemble. The final evaluation is performed on an unseen test set containing 6,323 slices. The soft-voting ensemble achieved the highest accuracy of 92.74%. These results show that the models can detect tumor and no-tumor slices, while the class imbalance remains an important challenge for tumor detection.