RETRACTED: Optimizing CNN performance in medical imaging with cross-modality pre-training: A study using MobileNetV3
Abstract
In this study, we perform an analysis of convolutional neural networks (CNNs) for transfer learning tasks, leveraging cross-organ and cross-modality transfer learning. Initially, we pre-train the models on an unrelated dataset of mammograms (5734 images), using this as an intermediate step before fine-tuning the models with traditional transfer learning (TL) on the ProstateX dataset, which serves as the target dataset. The CNN architecture evaluated was MobileNetV3. We assessed the model performance using the accuracy as the evaluation metric while also reporting segmentation metrics such as Dice coefficient. Our results demonstrate that MobileNetV3 outperformed VGG16 which is commonly used in previous studies in terms of accuracy, achieving an accuracy of 0.99 compared to VGG16. While segmentation specific metrics indicate room for improvement. This study highlights the advantage of cross-organ and cross-modality transfer learning in improving network performance, with MobileNetV3 showing significant potential in transfer learning tasks after pre-training on mammogram data. • Cross-organ, cross-modality transfer learning applied. • MobileNetV3 pre-trained on mammogram dataset. • Fine-tuned on ProstateX target dataset. • MobileNetV3 achieved 0.99 accuracy score. • Outperformed VGG16 in transfer learning tasks.