Development of a CAD system to distinguish between colon cancer and diverticulitis based on CT scans: a multicentre study

The objectives of this thesis are to develop a CAD (computer-aided detection/diagnosis) system capable of distinguishing between colon cancer and diverticulitis on computed tomography images of the colon (CTC), and to create an algorithm for the automatic segmentation of these volumes. Diverticulitis, or symptomatic diverticulosis, consists of the formation of growths in the colonic lumen which, when infected, inflame the abdominal wall; its incidence is increasing in Western countries. The symptoms of diverticulitis can overlap with those of colon cancer, the early diagnosis of which is crucial to improving the chances of recovery. Colonoscopy with biopsy is the gold standard for diagnosing colon cancer, but it is an invasive and demanding examination for the patient. To select high-risk candidates, “virtual colonoscopy” can be used an abdominal CTC that allows the radiologist to observe a three-dimensional reconstruction of the colon, segment the suspicious portions and decide whether to proceed with endoscopic examination. However, in challenging cases (typically elderly patients or those with marked wall thickening and luminal stenosis), the distinction between tumours and diverticula can be even more complex. This retrospective multicentre study involved 151 CT scans collected between January 2010 and September 2023 at two centres: 110 from the Luigi Sacco Hospital and 41 from the Spedali Civili Hospital. Two radiologists (with 20 and 2 years of experience) segmented 110 and 151 volumes, respectively. The set of 110 “expert database” volumes was used to train the models, while the remaining 41 CTCs formed the external validation dataset. Using PyRadiomics, in line with IBSI standards, 960 radiomic features were extracted per image. In order to assess their robustness in relation to the segmentations, the infraclass correlation coefficient (ICC) was calculated, but only on the features of the images whose segmentation masks were shown to have a dice similarity coefficient (DSC) greater than 0.6, i.e. 41 images. Only features with ICC>07 were taken into consideration, the number of which therefore stood at 279. The datasets were normalised using min-max normalisation and also discretised with ChiMerge discretisation. A 4-fold cross-validation was performed: in each fold, the minimum-Redundancy-Maximum-Relevance (mRMR) feature selection algorithm was applied and different types of classifiers were trained. The best models were selected based on sensitivity, balanced accuracy, F1 score, maximum specificity on the validation folds and on the same metrics evaluated only on challenging cases. Subsequently, the performance of these classifiers was measured on the external dataset and also on the same images used in the training phase but segmented by the radiologist in training, in order to evaluate the robustness of the models in terms of DSC. The best models (ANN, Lasso, SVM, KNN) achieved performance in terms of the above metrics greater than 0.8 on the validation folds and challenging cases. At the same time, nnU-Net, a self-configuring deep learning framework for biomedical image segmentation, was implemented. Trained on the “expert” database, the network achieved an average DSC of 0.60 in validation.