Hyperspectral pathology combined with uncertainty-calibrated deep learning for patient-level differentiation of adrenocortical tumors and pheochromocytomas.

Abstract

Background: Adrenocortical tumors (ACTs) and pheochromocytomas (PHEOs) can overlap morphologically. Hyperspectral imaging (HSI) provides label-free spectral-spatial information. In this study, reflectance HSI was conducted in the visible-to-near-infrared range (400-1000 nm), but conventional models incompletely capture long-range wavelength dependencies and correlated tissue patches.

Methods: We developed AdrenoMamba-MIL, a patient-level framework combining bidirectional spectral selective state-space modeling, multiscale spatial encoding, cross-gated fusion, attention multiple-instance learning, temperature scaling, and uncertainty estimation. This retrospective single-center study included 60 patients (30 ACTs and 30 PHEOs), partitioned at patient level into training (n=27), validation (n=9), internal test (n=9), and locked same-center held-out test (n=15) subsets. The primary endpoint was patient-level area under the receiver operating characteristic curve (AUC) in the locked test set.

Results: In the pooled test cohort (n=24), AdrenoMamba-MIL achieved an AUC of 0.948 (95% confidence interval [CI], 0.829-0.993), accuracy of 91.7%, sensitivity of 91.7%, specificity of 91.7%, and F1 score of 0.917. The locked-test AUC was 0.946 (95% CI, 0.804-0.997). Ablation analyses supported complementary contributions from the spectral, spatial, fusion, and patient-level aggregation components.

Conclusions: AdrenoMamba-MIL showed promising same-center performance for label-free ACT-PHEO differentiation. Independent external and prospective validation is required before clinical implementation.