Michele Maddalo

Background: To address the numerous unmeet clinical needs, in recent years several Machine Learning models applied to medical images and clinical data have been introduced and developed. Even when they achieve encouraging results, they lack evolutionary progression, thus perpetuating their status as autonomous entities. We postulated that different algorithms which have been proposed in the literature to address the same diagnostic task, can be aggregated to enhance classification performance. We suggested a proof of concept to define an ensemble approach useful for integrating different algorithms proposed to solve the same clinical task. Methods: The proposed approach was developed starting from a public database consisting of radiomic features extracted from CT images relating to 535 patients suffering from lung cancer. Seven algorithms were trained independently by participants in the AI4MP working group on Artificial Intelligence of the Italian Association of Physics in Medicine to discriminate metastatic from non-metastatic patients. The classification scores generated by these algorithms are used to train SVM classifier. The Explainable Artificial Intelligence approach is applied to the final model. The ensemble model was validated following an 80–20 hold-out and leave-one-out scheme on the training set. Results: Compared to individual algorithms, a more accurate result was achieved. On the independent test the ensemble model achieved an accuracy of 0.78, a F1-score of 0.57 and a log-loss of 0.49. Shapley values representing the contribution of each algorithm to the final classification result of the ensemble model were calculated. This information represents an added value for the end user useful for evaluating the appropriateness of the classification result on a particular case. It also allows us to evaluate on a global level which methodological approaches of the individual algorithms are likely to have the most impact. Conclusion: Our proposal represents an innovative approach useful for integrating different algorithms that populate the literature and which lays the foundations for future evaluations in broader application scenarios.

(1) Background and (2) Methods: In this retrospective, observational, monocentric study, we selected a cohort of eighty-five patients (age range 38–87 years old, 51 men), enrolled between January 2014 and December 2020, with a newly diagnosed renal mass smaller than 4 cm (SRM) that later underwent nephrectomy surgery (partial or total) or tumorectomy with an associated histopatological study of the lesion. The radiomic features (RFs) of eighty-five SRMs were extracted from abdominal CTs bought in the portal venous phase using three different CT scanners. Lesions were manually segmented by an abdominal radiologist. Image analysis was performed with the Pyradiomic library of 3D-Slicer. A total of 108 RFs were included for each volume. A machine learning model based on radiomic features was developed to distinguish between benign and malignant small renal masses. The pipeline included redundant RFs elimination, RFs standardization, dataset balancing, exclusion of non-reproducible RFs, feature selection (FS), model training, model tuning and validation of unseen data. (3) Results: The study population was composed of fifty-one RCCs and thirty-four benign lesions (twenty-five oncocytomas, seven lipid-poor angiomyolipomas and two renal leiomyomas). The final radiomic signature included 10 RFs. The average performance of the model on unseen data was 0.79 ± 0.12 for ROC-AUC, 0.73 ± 0.12 for accuracy, 0.78 ± 0.19 for sensitivity and 0.63 ± 0.15 for specificity. (4) Conclusions: Using a robust pipeline, we found that the developed RFs signature is capable of distinguishing RCCs from benign renal tumors.

Objectives To test reproducibility and predictive value of a simplified score for assessment of extraprostatic tumor extension (sEPE grade). Methods Sixty-five patients (mean age ± SD, 67 years ± 6.3) treated with radical prostatectomy for prostate cancer who underwent 1.5-Tesla multiparametric magnetic resonance imaging (mpMRI) 6 months before surgery were enrolled. sEPE grade was derived from mpMRI metrics: curvilinear contact length > 15 mm (CCL) and capsular bulging/irregularity. The diameter of the index lesion (dIL) was also measured. Evaluations were independently performed by seven radiologists, and inter-reader agreement was tested by weighted Cohen K coefficient. A nested (two levels) Monte Carlo cross-validation was used. The best cut-off value for dIL was selected by means of the Youden J index to classify values into a binary variable termed dIL*. Logistic regression models based on sEPE grade, dIL, and clinical scores were developed to predict pathologic EPE. Results on validation set were assessed by the main metrics of the receiver operating characteristics curve (ROC) and by decision curve analysis (DCA). Based on our findings, we defined and tested an alternative sEPE grade formulation. Results Pathologic EPE was found in 31/65 (48%) patients. Average κ_w was 0.65 (95% CI 0.51–0.79), 0.66 (95% CI 0.48–0.84), 0.67 (95% CI 0.50–0.84), and 0.43 (95% CI 0.22–0.63) for sEPE grading, CLL ≥ 15 mm, dIL*, and capsular bulging/irregularity, respectively. The highest diagnostic yield in predicting EPE was obtained by combining both sEPE grade and dIL*(ROC-AUC 0.81). Conclusions sEPE grade is reproducible and when combined with the dIL* accurately predicts extraprostatic tumor extension. Key Points • Simple and reproducible mpMRI semi-quantitative scoring system for extraprostatic tumor extension. • sEPE grade accurately predicts extraprostatic tumor extension regardless of reader expertise. • Accurate pre-operative staging and risk stratification for optimized patient management.

Purpose: To test radiomics for prognostication of intrahepatic mass-forming cholangiocarcinoma (IMCC) and to develop a comprehensive risk model. Methods: Histologically proven IMCC (representing the full range of stages) were retrospectively analyzed by volume segmentation on baseline hepatic venous phase computed tomography (CT), by two readers with different experience (R1 and R2). Morphological CT features included: tumor size, hepatic satellite lesions, lymph node and distant metastases. Radiomic features (RF) were compared across CT protocols and readers. Univariate analysis against overall survival (OS) warranted ranking and selection of RF into radiomic signature (RSign), which was dichotomized into high and low-risk strata (RSign*). Models without and with RSign* (Model 1 and 2, respectively) were compared. Results: Among 78 patients (median follow-up 262 days, IQR 73–957), 62/78 (79%) died during the study period, 46/78 (59%) died within 1 year. Up to 10% RF showed variability across CT protocols; 37/108 (34%) RF showed variability due to manual segmentation. RSign stratified OS (univariate: HR 1.37 for R1, HR 1.28 for R2), RSign* was different between readers (R1 0.39; R2 0.57). Model 1 showed AUC 0.71, which increased in Model 2: AUC 0.81 (p < 0.001) and AIC 89 for R1, AUC 0.81 (p = 0.001) and AIC 90.2 for R2. Conclusion: The use of RF into a unified RSign score stratified OS in patients with IMCC. Dichotomized RSign* classified survival strata, its inclusion in risk models showed adjunct yield. The cut-off value of RSign* was different between readers, suggesting that the use of reference values is hampered by interobserver variability.