A hybrid feature pyramid network and Efficient Net-B0-based GIST detection and segmentation from fused CT-PET image

Abstract

Gastrointestinal stromal tumour (GIST) is the type of tumour that develops in the gastrointestinal tract and disrupts the digestive system. The two most crucial factors for GIST detection and monitoring are structural and functional information, either one of which can be determined by a single imaging modality. Computed tomography (CT) and positron emission tomography (PET) are the two non-invasive imaging techniques that are frequently employed for this purpose. Effective clinical management dependent on precise diagnosis, which centres on accurate tumour detection and segmentation, is the major concern with manual processes. To over this problem, we proposed a deep learning framework in two stages: in the first stage, we use Dense Net and VGG model for CT and PET image fusion, and in the second stage, for effective tumour segmentation we use feature pyramid network (FPN) for automatic and accurate tumour detection. The advantage of using Dense Net and VGG Net for medical image fusion is their ability to extract features at different scales and resolutions. This allows for a more comprehensive representation of the input images, which can be beneficial from medical image fusion. The fused image is segmented using FPN semantic segmentation model with Efficient Net-B0 as encoder; finally the proposed model performance was evaluated through qualitative and quantitative metrics. The average mean squared error value is 63.27, the average peak signal-to-noise ratio value is 30.154 and the average Structural Similarity Index Measure value is 0.9647. The proposed FPN with Efficient Net-B0 segmentation produced an intersection over union of 0.8521 and pixel accuracy 96.09, which are comparatively higher than the other existing models. The proposed deep learning framework outperformed well in automatically and accurately segmenting GIST, assisting doctors in effective clinical management.