Calibration and Cross-Dataset Behavior in Malaria Microscopy with Parameter-Efficient Fine-Tuning: A Matched LoRA–PiSSA Evaluation

Background/Objectives: Reliable evaluation of deep-learning methods for malaria microscopy requires assessing both classification performance and the quality of predicted infection probabilities across datasets. We compared low-rank adaptation (LoRA) and principal singular values and singular vectors adaptation (PiSSA) under matched parameter-efficient fine-tuning (PEFT). Methods: An ImageNet-pretrained vision transformer (ViT-B/16) was adapted at ranks 2, 4, 8, and 16 using 27,558 National Institutes of Health (NIH) malaria cell images. One NIH grouped partition was held fixed across paired training seeds 42, 123, and 456. Models were evaluated on the NIH test split and, without target-domain adaptation, on BBBC041 P. vivax cell crops. The analysis was also replicated in invasive ductal carcinoma (IDC) histopathology. External probability and threshold-transfer analyses were post hoc. Results: Across LoRA and PiSSA configurations, NIH area under the receiver operating characteristic curve (ROC–AUC) ranged from 0.9960 to 0.9970. PiSSA had lower rank-level mean Brier scores and expected calibration error (ECE) at every evaluated rank in NIH and IDC, although paired-run differences varied. On BBBC041, PiSSA-r16 achieved the highest PEFT ROC–AUC (0.8451) and area under the precision–recall curve (PR–AUC, 0.1588), while fully trained reference models performed better. Across all 12 matched external comparisons, PiSSA had lower overall Brier scores and ECE but higher infected-cell Brier scores. PiSSA had lower sensitivity at NIH-derived thresholds in 11 of 12 comparisons. All PEFT models overpredicted infection relative to the observed 2.70% prevalence. Conclusions: Nearly identical internal discrimination concealed differences in external probability behavior and class-specific errors. PiSSA reduced overprediction, but neither adapter achieved adequate external calibration.

Authors

Institutions

Publication Details

Journal
Diagnostics
Published
2026-11-12
DOI
https://doi.org/10.3390/diagnostics16223749
Primary Topic
Digital Imaging for Blood Diseases
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Calibration and Cross-Dataset Behavior in Malaria Microscopy with Parameter-Efficient Fine-Tuning: A Matched LoRA–PiSSA Evaluation

Rajvardhan Patil, Sorio Boit, Tiehang Duan
Diagnostics
Digital Imaging for Blood Diseases
article

Calibration and Cross-Dataset Behavior in Malaria Microscopy with Parameter-Efficient Fine-Tuning: A Matched LoRA–PiSSA Evaluation

Rajvardhan Patil, Sorio Boit, Tiehang Duan
article en

Abstract

Background/Objectives: Reliable evaluation of deep-learning methods for malaria microscopy requires assessing both classification performance and the quality of predicted infection probabilities across datasets. We compared low-rank adaptation (LoRA) and principal singular values and singular vectors adaptation (PiSSA) under matched parameter-efficient fine-tuning (PEFT). Methods: An ImageNet-pretrained vision transformer (ViT-B/16) was adapted at ranks 2, 4, 8, and 16 using 27,558 National Institutes of Health (NIH) malaria cell images. One NIH grouped partition was held fixed across paired training seeds 42, 123, and 456. Models were evaluated on the NIH test split and, without target-domain adaptation, on BBBC041 P. vivax cell crops. The analysis was also replicated in invasive ductal carcinoma (IDC) histopathology. External probability and threshold-transfer analyses were post hoc. Results: Across LoRA and PiSSA configurations, NIH area under the receiver operating characteristic curve (ROC–AUC) ranged from 0.9960 to 0.9970. PiSSA had lower rank-level mean Brier scores and expected calibration error (ECE) at every evaluated rank in NIH and IDC, although paired-run differences varied. On BBBC041, PiSSA-r16 achieved the highest PEFT ROC–AUC (0.8451) and area under the precision–recall curve (PR–AUC, 0.1588), while fully trained reference models performed better. Across all 12 matched external comparisons, PiSSA had lower overall Brier scores and ECE but higher infected-cell Brier scores. PiSSA had lower sensitivity at NIH-derived thresholds in 11 of 12 comparisons. All PEFT models overpredicted infection relative to the observed 2.70% prevalence. Conclusions: Nearly identical internal discrimination concealed differences in external probability behavior and class-specific errors. PiSSA reduced overprediction, but neither adapter achieved adequate external calibration.

DiagnosticsVol. 16(22)
Grand Valley State University (US)
Good health and well-being
Digital Imaging for Blood Diseases
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