Training-Volume Analysis, Computational Efficiency, and Green AI in Rotator Cuff Calcific Tendinopathy...

To evaluate how training data volume influences diagnostic performance, external validation, robustness to training-subset composition, and computational efficiency in a VGG19-based model for rotator cuff calcific tendinopathy (RCCT) classification on shoulder radiographs (XRs), using training data from a single development institution. This retrospective two-institution study included 5056 shoulder XRs. A Visual Geometry Group 19 (VGG19)-based pipeline was trained on four cumulative fractions of a 4268-XRs development cohort: 25% (M25), 50% (M50), 75% (M75), and 100% (M100), using stratified 5-fold cross-validation. To assess robustness to training-subset composition, three nested cumulative sampling realizations were generated for M25, M50, and M75, while M100 represented the fixed full-data reference. Models were evaluated on an internal test cohort ( $n=480$ ) and an external validation cohort ( $n=308$ ). Areas under the receiver operating characteristic curves (AUCs) were compared using correlated DeLong tests with Holm adjustment. M75 and M100 were additionally assessed using two one-sided tests (TOST) with a $\pm 0.02$ AUC equivalence margin. Training-volume robustness was assessed across repeated cumulative subset realizations. Grad-CAM localization was quantitatively evaluated against expert-defined lesion regions using Pointing Game, localization area under the precision-recall curve (AUPRC), and Energy-Based Pointing Game (EBPG), with Score-CAM used for qualitative corroboration. In the primary sampling realization, internal AUC increased from 0.869 with M25 to 0.951 with M100. External AUC increased from 0.835 (M25) to 0.903 (M50), 0.945 (M75), and 0.953 (M100). External improvements were significant from M25 to M50 and from M50 to M75 (both Holm-adjusted $P<.001$ ), whereas the M75-to-M100 increment was small ( $\Delta$ AUC=0.008; $P=.161$ ). M75 and M100 were externally equivalent within the $\pm 0.02$ AUC margin (TOST $P=.016$ ). Across three independently sampled M75 subsets, external AUC remained highly stable (0.945-0.948; mean ± standard deviation, 0.946 ± 0.001), and all three M75 realizations were equivalent to M100. Compared with M100, M75 required 11.4% less training time and 15.4% less measured energy. In the primary seed-42 realization, quantitative Grad-CAM localization generally improved with increasing training volume: M100 achieved the highest Pointing Game and localization AUPRC in both cohorts, whereas EBPG peaked at M75. Increasing training volume improved model discrimination, but external gains showed marked diminishing returns beyond M75 in this VGG19-based RCCT classification task. Near-maximal external performance at M75 was reproducible across independent training-subset realizations and statistically equivalent to full-data training within the evaluated margin, supporting training-volume analysis as a strategy for balancing performance and computational resources within this study setting. In this VGG19-based RCCT classification task using training data from a single development institution, external discrimination showed marked diminishing returns beyond 75% training volume, with highly reproducible M75 performance across independent subset realizations and statistical equivalence to full-data training within a $\pm 0.02$ AUC margin.