Multi-Scale Network for Thoracic Organs Segmentation

Medical Imaging Segmentation is an essential technique for modern medical applications. It the foundation of many aspects clinical diagnosis, oncology, and computer-integrated surgical intervention. Although significant successes have been achieved in segmentation images, DL (deep learning) approaches. Manual delineation OARs (organs at risk) vastly dominant but it prone to errors given complex irregularities shape, low texture diversity between tissues adjacent blood area, patient-wide location organisms, weak soft tissue contrast across organs CT images. Till now several models implemented on multi not caters problem imbalanced classes some relatively small pixels as compared others. To segment thoracic we proposed model based encoder-decoder approach using transfer learning with efficientnetB7 model. We built a fully connected CNN (Convolutional Neural network) having 5 layers encoding decoding specifically tackle imbalance class accurate way OARs. Proposed methodology achieves 0.93405 IOU score, 0.95138 F1 score class-wise dice esophagus 0.92466, trachea 0.94257, heart 0.95038, aorta 0.9351 background 0.99891. The results showed that our framework can be segmented accurately.