A Novel Neural Network for Joint Lesion Segmentation and Confidence Score Generation from PET Image

Lesions segmentation from PET images is considered very high challenging task compared to the anatomical organ delineation regarding irregular and/or unpredictable shape/morphology of lesions. Moreover, lesion segmentation from PET images alone would add to the complexity of the problem owing to the poor spatial resolution and high levels of noise. Thus, dedicated/optimized segmentation models should be developed for identification and delineation of malignant lesions from PET images. To this end, this work set out to propose a novel solution for this challenge. Moreover, the focus of this study is to introduce an automated model assigning a confidence score to the resulting segmentation in order to indicate to what extend specialists could trust the outcomes. This would greatly reduce the workload and gross errors in clinical practice. To this end, a GAN network was developed in which a discriminator repeatedly evaluates the accuracy of the estimated lesion segmentation. This module is trained to identify the accurate estimations. This module sends feedback to the primary segmentation network to improve the overall segmentation accuracy as well as providing a confidence score which indicates the accuracy of the final segmentation. Regarding the quantitative analysis of the proposed network, the incorporation of the confidence score estimator improved the segmentation accuracy of the model from 85.9 % (without) to 86.8% (with the confidence module). Moreover, the confidence module enabled to estimate the accuracy of the resulting segmentation with a mean absolute error (MAE) of 0.084 compared to the original model with MAE of 0.159. The proposed confidence score estimator would minimize the incidence of gross errors in clinical practice as well as reducing the workload for verification of the resulting segmentations.