Addressing the Imbalanced Class Distribution in Fatty Liver Detection in CT Images Using Transfer Learning

Addressing the common challenges posed by limitations in medical images, our study involved the collection of CT images from the abdominal area of patients. We meticulously selected the appropriate shot for diagnosing fatty liver disease and employed the appropriate data augmentation techniques suitable for medical images to overcome the scarcity of positive samples. Three pre-trained models, namely InceptionResNetV2, EfficientNetB4, and EfficientNetV2S, were used in the study to focus on the use of transfer learning; the performance of these models was evaluated both before and after applying data augmentation techniques. After obtaining the results, the EfficientNetV2S model produced promising results for the fatty liver disease classification. F1-score, Test Accuracy, and Test Loss on the balanced dataset were 0.86, 0.93, and 0.19, respectively.