Abstract
Cervical cancer progresses through a precancerous stage that, if detected early, can be treated before invasive carcinoma develops.
However, most deep learning approaches to colposcopy-based screening frame the problem as binary: normal-versus-abnormal or cancer-versus non-cancer classification, overlooking the precancerous class that is most actionable for prevention. Whether task framing affects the accurate detection of precancerous lesions remains underexplored. We present a hierarchical task framing for cervical lesion grading that decomposes the problem into a cascade of clinically motivated binary decisions. We compare this against conventional 4-class and 3-class task framings to investigate which best supports precancerous detection. Experiments are conducted on a Nigerian colposcopy screening dataset comprising 2,835 images from 282 patients, labelled across four clinically grounded severity grades: Normal, Borderline, Low-grade Precancerous, and High-grade Precancerous. Two deep learning architectures are evaluated across all
three task framings under strict patient-level splits. Results show that under oracle stage-wise evaluation, hierarchical task framing consistently achieves higher macro-recall across all stages, with EfficientNet-B3 reaching a best stage macro-recall of 0.557 ± 0.022, versus 0.227 ± 0.020 and 0.366 ± 0.023 for the 4-class and 3-class task framings respectively. However, end-to-end cascade evaluation of the hierarchical framing under real routing yields a macro-recall of 0.372 ± 0.104, comparable to the 3-class flat baseline, suggesting that the benefit of hierarchical framing depends on upstream stage accuracy and motivating future work on error-robust cascade designs.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MIRASOL_024.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/MIRASOL_024_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=pk1Uouqlbo
BibTex
@InProceedings{AdeAra_Hierarchical_MICCAISAT2026,
author = { Adebesin, Aramide AND Ugbo, Ikechukwu AND Ogundeji, Olawale AND Shiaki, Nunsi AND Adegoke, Solomon AND Shiaki, Tsintop AND Gbayisomore, Tolulope Judah AND Iorumbur, Aondona Moses AND Raymond, Confidence},
title = { { Hierarchical Task Framing for Prevention-Oriented Cervical Lesion Detection Using Nigerian Colposcopy Images } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17264},
month = {pending},
page = {pending}
}