A heterogeneous deep-learning ensemble for assistive 11-class classification of kidney diseases using transmission electron microscopy

Abstract Transmission electron microscopy (TEM) aids glomerular disease diagnosis. However, single-image classification under patient-independent evaluation remains uncertain. We classified 11 kidney diseases using 60,509 quality-controlled TEM images from 8,250 patients, partitioned by patient into training (5,940 patients), meta-calibration (Meta-Cal; 661), and held-out test (1,649) sets. Five backbones (EfficientNet-B4, ViT-B/16, Swin-B, ConvNeXt-B, and DeiT-B) were trained using patient-grouped nested folds. Training-set out-of-fold predictions were used to fit weighted soft voting and logistic-regression stacking with eightfold test-time augmentation. Meta-calibration compared these two prespecified candidates and supported calibration and referral-rule assessment. Single-image accuracy, macro-F1 score, balanced accuracy, Matthews correlation coefficient, and macro-AUC (area under the receiver operating characteristic curve) were 0.444, 0.361 (bootstrap mean: 0.360; 95% confidence interval: 0.344–0.379), 0.354, 0.345, and 0.793, respectively. Averaging probabilities across images of 1,642 single-diagnosis patients yielded an accuracy of 0.610 and macro-F1 score of 0.507; 7 patients with discordant labels were excluded only from this aggregation. These results were not comparable with image-level-split estimates, including our original submission. Temperature scaling reduced expected calibration error from 0.059 to 0.028. The prespecified referral target (Meta-Cal autonomous accuracy ≥ 0.90) was not reached, and therefore no operating threshold was selected. The results provide a patient-independent benchmark, not evidence for autonomous diagnosis.

Authors

Institutions

Publication Details

Journal
Scientific Reports
Published
2026-11-16
DOI
https://doi.org/10.1038/s41598-026-78000-8
Primary Topic
AI in cancer detection
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

A heterogeneous deep-learning ensemble for assistive 11-class classification of kidney diseases using transmission electron microscopy

Junpu Wang, Jun Li, Huiling Wang, Weiyuan Wang et al.
4 links
Scientific Reports
AI in cancer detection
article

A heterogeneous deep-learning ensemble for assistive 11-class classification of kidney diseases using transmission electron microscopy

Junpu Wang, Jun Li, Huiling Wang, Weiyuan Wang, Haixin Wang, Liu Liu, Ziyu Liu, Xiaoyang Li, Qiqi Zhang, Jie Liu
article en

Abstract

Abstract Transmission electron microscopy (TEM) aids glomerular disease diagnosis. However, single-image classification under patient-independent evaluation remains uncertain. We classified 11 kidney diseases using 60,509 quality-controlled TEM images from 8,250 patients, partitioned by patient into training (5,940 patients), meta-calibration (Meta-Cal; 661), and held-out test (1,649) sets. Five backbones (EfficientNet-B4, ViT-B/16, Swin-B, ConvNeXt-B, and DeiT-B) were trained using patient-grouped nested folds. Training-set out-of-fold predictions were used to fit weighted soft voting and logistic-regression stacking with eightfold test-time augmentation. Meta-calibration compared these two prespecified candidates and supported calibration and referral-rule assessment. Single-image accuracy, macro-F1 score, balanced accuracy, Matthews correlation coefficient, and macro-AUC (area under the receiver operating characteristic curve) were 0.444, 0.361 (bootstrap mean: 0.360; 95% confidence interval: 0.344–0.379), 0.354, 0.345, and 0.793, respectively. Averaging probabilities across images of 1,642 single-diagnosis patients yielded an accuracy of 0.610 and macro-F1 score of 0.507; 7 patients with discordant labels were excluded only from this aggregation. These results were not comparable with image-level-split estimates, including our original submission. Temperature scaling reduced expected calibration error from 0.059 to 0.028. The prespecified referral target (Meta-Cal autonomous accuracy ≥ 0.90) was not reached, and therefore no operating threshold was selected. The results provide a patient-independent benchmark, not evidence for autonomous diagnosis.

Scientific Reports
Central South University (CN), Xiangya Hospital Central South University (CN)
Central South University
Good health and well-being
AI in cancer detection
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon