Deep Learning-Based Diagnosis of Disease Activity in Patients with Graves’ Orbitopathy Using Orbital SPECT/CT

Abstract

Purpose Orbital 99mTc-DTPA SPECT/CT is an important new method for the assessment of inflammatory activity in patients with Graves' Orbitopathy (GO), but it consumes a heavy workload for physicians for interpretation. We aim to propose an automated method, called GO-Net, to detect the activity of GO to assist physicians for diagnosis. Materials and methods GO-Net had two stages: a semantic V-Net segmentation network (SV-Net) to extract extraocular muscles (EOMs) on orbital CT images; a three-channel convolutional neural network (CNN), including SPECT/CT images and segmentation results, to classify inflammatory activity. Manual corrections were applied when the segmentation results were not accurate. A total of 956 eyes from 478 patients with GO (active: 475; inactive: 481) from Xiangya Hospital of Central South University were enrolled. For the segmentation, five-fold cross-validation with 194 eyes were used for training and internal validation. For the classification, 80% of eyes were trained and internally validated by five-fold cross-validation, and 20% of eyes were used for testing. The contours of the EOMs were drawn manually by an experienced physicians and used as the ground truth. The criteria for the diagnosis of GO activity were determined by the physician through the clinical activity score(CAS) and 99mTc-DTPA uptake. Results Our GO-Net method achieved an accuracy of 84.25%, a precision of 83.35%, a sensitivity of 84.63%, a specificity of 83.87%, an F1 score of 0.83, and an area under the receiver (AUC) of 0.89. For EOMs segmentation, our segmentation model achieved a mean intersection over union (IOU) of 0.82. Contours of EOMs in 47 eyes (4.91%) were manually corrected and the average correction time was 5 mins for each eye. Conclusion Our proposed Go-Net model could accurately detect GO activity, which has great potential for the diagnosis of GO.

Computed tomography‐based artificial intelligence in lung disease—Chronic obstructive pulmonary disease

“…Linardators, Papastefanopoulos, and Kotsiantis have reviewed multiple machine learning interpretability methods in the field of XAI, including guided saliency, class activation maps (CAM), Grad‐CAM++, and so on 44 . Yao et al 45 applied Grad‐CAM on interpreting the CT‐based models, demonstrating a way of dealing with the “black box” issue.…”