Abstract
Cervical cancer remains the leading cause of cancer-related death among women in sub-Saharan Africa, where limited access to expert colposcopists delays diagnosis and treatment. Deep learning models trained on colposcopy images have reported accuracies approaching or exceeding colposcopist-level discrimination, but colposcopy datasets typically contain multiple correlated images per woman (different magnifications and pre-/post-acetic-acid views), and published evaluations rarely state whether train/test splits were constructed at the patient or image level. We built a colposcopy classification pipeline combining unsupervised, mask-free cervix localization with four ImageNet-pretrained backbones (ResNet18, MobileNetV3-Small, EfficientNet-B0, DenseNet121) and a class-weighted Dirichlet evidential head, and evaluated it on a public dataset of 3,356 images from 332 women screened at a Nigerian teaching hospital under two splitting protocols on the same data: a naive image-level split and a patient-level split in which no woman’s images cross between training, validation, and test. The image-level split reproduced the high performance reported in prior colposcopy-classification work (best baseline AUC 0.914, proposed model AUC 0.934); under the patient-level split, every model’s performance collapsed to near chance (AUC 0.495–0.545), with the proposed evidential model achieving the highest AUC (0.545) and specificity (0.633) among the five models compared, but margins this small are not distinguishable given a 50-patient test set. An ablation study confirms that the class-weighting and partial-fine-tuning components still serve their intended purpose, removing either causes the model to collapse toward majority-class prediction, even though the resulting genuine discriminative signal remains weak. These results are a direct, matched-pipeline demonstration that naive image-level evaluation can substantially overstate colposcopy-classification performance on small-cohort datasets, and we argue that patient-level splitting should be treated as a mandatory reporting requirement in this literature.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/CAPI_WOMEN_013.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=oE0PLD1sId
BibTex
@InProceedings{LakAli_PatientLevel_MICCAISAT2026,
author = { Lakyiere, Alice Bagyiereyele AND Adusei, Richard AND Owusu-Agyemang, Kwabena AND Gyening, Rose-Mary Owusuaa Mensah},
title = { { Patient-Level Data Leakage in Colposcopy Classification: A Cautionary Study with Evidential Uncertainty Quantification for Cervical Cancer Screening } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17256},
month = {pending},
page = {pending}
}