Phd in EXPERIMENTAL MEDICINE

Thesis title: Radiomica e Intelligenza Artificiale per la Caratterizzazione di Lesioni Renali Solide in Tomografia Computerizzata

BACKGROUND AND RATIONALE: The incidence of renal tumors is steadily increasing, largely due to the rise in incidental diagnoses related to the widespread use of abdominal imaging. However, a non-negligible proportion of resected renal lesions prove to be histologically benign or of low aggressiveness, with a consequent risk of surgical overtreatment. The main challenges concern the differentiation between renal cell carcinoma (RCC), including its histotypes of clear cell RCC (ccRCC) and chromophobe RCC (crRCC), and benign tumors such as angiomyolipoma (AML) and oncocytoma, which often share many imaging and histopathological features. In this context, radiomics applied to computed tomography (CT) enables the quantification of intra-tumoral heterogeneity, a potential biomarker of biological aggressiveness, but its clinical use is still limited and poorly standardized. The aim of the present study was to assess the diagnostic performance of selected radiomic features extracted from contrast-enhanced CT images, subsequently integrated into a machine learning (ML) decision-tree model, to distinguish malignant from benign renal lesions and, in secondary analyses, to differentiate RCC from AML and chRCC from oncocytoma. MATERIALS AND METHODS: A retrospective study was conducted on consecutive patients evaluated between May 2022 and January 2025, who underwent multiphasic contrast-enhanced abdominal CT for solid renal lesions and had a definitive histological diagnosis after surgery or biopsy. Cases without pre-treatment CT, without histologic confirmation, with predominantly cystic lesions, or with non-diagnostic images were excluded. All examinations were performed on a dual-source multidetector CT scanner (384 slices, Siemens SOMATOM Force) using a standardized protocol: breath-hold acquisitions, 120 kVp, 200–650 mAs with automatic modulation, 4–5 mm slice thickness, 512×512 matrix, with unenhanced, arterial, venous/portal, and delayed phases after intravenous administration of iodinated contrast medium (1,5 ml/kg). For the radiomic analysis, only the venous/portal phase (75 s) was used, as considered more homogeneous for assessment of the renal parenchyma. Lesions were initially assessed morphologically (maximum axial diameter) and subsequently manually segmented on the three planes (axial, coronal, sagittal) using 3D Slicer, generating three-dimensional volumes of interest (VOIs), with a threshold of –50 HU applied to exclude gas and peripheral adipose tissue. On the filtered images, a filtration–histogram technique was applied using three Spatial Scaling Factors (SSF = 2 mm, 4 mm, 6 mm): the fine filter (SSF = 2) enhances parenchymal micro-heterogeneity, while the medium and coarse filters (SSF = 4, 6) emphasize coarser and vascular components. From each VOI, four specific radiomic features (entropy, kurtosis, MPP, skewness) were extracted at each filtering level using PyRadiomics. Features were analyzed using binary logistic regression to identify independent predictors of malignancy, and receiver operating characteristic (ROC) curves (AUC, sensitivity, specificity) were calculated to evaluate their diagnostic performance. Features that proved significant were finally integrated into a decision-tree ML algorithm (J48, WEKA software) to obtain an interpretable classification model capable of discriminating among ccRCC, chRCC, AML, and oncocytoma. RESULTS: The study population consisted of 112 patients (mean age 61 ± 13 years; 56% male, 44% female), of whom 65% had malignant and 35% benign lesions; malignant lesions were on average larger than benign ones (4,5 ± 3,5 vs 3,4 ± 2,4 cm; p < 0,001). On logistic regression analysis comparing malignant with benign lesions, only entropy and skewness with fine filtering (SSF = 2) were significantly associated with malignancy: entropy (OR = 13,72; p = 0,021) and skewness (OR = 2,06; p = 0,033) increased the probability of classifying a lesion as malignant, whereas kurtosis and MPP did not show significant associations. Malignant lesions exhibited higher mean entropy values (SSF = 2) than benign ones (5,49 ± 0,31 vs 5,38 ± 0,26; p = 0,036) and higher (less negative) skewness values (–0,09 ± 0,30 vs –0,28 ± 0,52; p = 0,012), indicating greater structural complexity and a more asymmetric pixel gray-level distribution. In the discrimination between malignant and benign lesions, entropy (SSF = 2; cut-off > 5,63) yielded an AUC of 0,581, with 86% specificity and 32% sensitivity, whereas skewness (SSF = 2; cut-off > 0,03) yielded an AUC of 0,622, with 81% specificity and 53% sensitivity. In the RCC vs AML sub-analysis, an increase in entropy (SSF = 2) was associated with a higher probability of malignancy (OR = 1,03; p = 0,048), with higher mean values in RCC compared with AML (5,49 ± 0,31 vs 5,31 ± 0,27; p = 0,009), and ROC performance characterized by an AUC of 0,659 (cut-off > 5,68), 98% specificity, and 34% sensitivity. In the distinction between chRCC and oncocytoma, entropy (SSF = 2) showed the opposite effect, with significantly higher mean values in oncocytoma than in chRCC (5,43 ± 0,26 vs 5,34 ± 0,19; p = 0,026), an AUC of 0,631 (cut-off > 5,58), 92% specificity, and 35% sensitivity. The J48 ML algorithm selected a final model consisting of three main features: entropy at SSF = 2 (2_entropy), entropy at SSF = 4 (4_entropy), and skewness at SSF = 2 (2_skewness). The root node was represented by 2_entropy (primary threshold 5.12), allowing an initial separation between low- and high-heterogeneity lesions; when 2_entropy ≤ 5,12, 4_entropy discriminated between AML and ccRCC, whereas for 2_entropy > 5,12, 2_skewness directed classification either towards ccRCC or towards a further node based on 2_entropy (threshold 5,27) to distinguish chRCC from oncocytoma. Overall, the model provides a schematic decision pathway for classification of the main histotypes considered. DISCUSSION AND CONCLUSIONS: The results confirm that specific radiomic measures of heterogeneity and pixel value distribution on CT images, particularly entropy and skewness with fine spatial filtering (SSF = 2), are significantly associated with malignancy in solid renal lesions and provide additional value in distinguishing RCC from AML, as well as chRCC from oncocytoma. Although the AUC values obtained are overall moderate, the high specificity observed for entropy (and, to a lesser extent, skewness) suggests a potential role for these parameters as “rule-in” markers, useful for strengthening the suspicion of malignancy when such values are elevated. Integration of the significant features into an interpretable ML model (J48 decision tree) makes it possible to transform abstract radiomic measures into explicit decision rules that are readily understandable to radiologists, thereby improving transparency and potential clinical acceptability. In selected clinical scenarios, such as indeterminate incidentally detected renal lesions or patients in whom the decision between biopsy, surgery, or active surveillance is challenging, a high-specificity radiomic model could function as a quantitative “second reader,” complementing conventional morphological assessment and contributing to the reduction of unnecessary surgical resections of benign lesions.