Title: Leakage-safe case-level evaluation of ResNet50V2 for benign-malignant liver tumour classification on CT slices

Authors: B. Smitha; R.S. Vinod Kumar; S.S. Kumar

Addresses: Department of Electronics and Communication Engineering, Noorul Islam Centre for Higher Education, Kumaracoil, Tamil Nadu, 629180, India ' Department of Electronics and Communication Engineering, Noorul Islam Centre for Higher Education, Kumaracoil, Tamil Nadu, 629180, India ' Department of Electronics and Instrumentation Engineering, Noorul Islam Centre for Higher Education, Kumaracoil, Tamil Nadu, 629180, India

Abstract: Binary (benign vs. malignant) liver tumour classification from computed tomography (CT) slices is often evaluated using slice-level splits that can introduce patient-level leakage. We present a leakage-safe, case-level evaluation of ResNet50V2 using stratified 5-fold cross-validation repeated over three random seeds. Slice probabilities were aggregated per Case_ID using mean probability (primary) and vote rate (secondary), with decision thresholds derived only from validation folds. Two preprocessing variants were assessed (BASIC and HE_GAUSS_CONTRAST). Across 15 runs, pooled case-level discrimination was AUC = 0.765 ± 0.10, with BASIC showing higher accuracy and lower variability than HE_GAUSS_CONTRAST. Calibration was assessed using expected calibration error, Brier score, and negative log-likelihood. The proposed framework provides a transparent baseline for small cohorts and supports future external validation.

Keywords: leakage-safe evaluation; case-level cross-validation; liver tumour classification; computed tomography; CT; ResNet50V2; slice aggregation; calibration.

DOI: 10.1504/IJBET.2026.156063

International Journal of Biomedical Engineering and Technology, 2026 Vol.51 No.2, pp.97 - 126

Received: 17 Sep 2025
Accepted: 16 Feb 2026
Published online: 02 Sep 2026 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article