HemoPrint: A Trustworthy Framework for Non-Invasive Blood Group Prediction
编号:215
访问权限:仅限参会人
更新:2026-11-15 14:06:00 浏览:11次
Online
摘要
Before a transfusion, an operation or many obstetric procedures, a patient's blood group has to be known, and finding it out still means drawing blood, applying antisera and waiting on a laboratory. We asked how far fingerprint images alone can go towards predicting the eight ABO/Rh groups (A+, A−, B+, B−, AB+, AB−, O+, O−), and built a framework, HemoPrint, to test this. A validation gate first rejects any input that is not a fingerprint. Accepted images are enhanced with contrast-limited adaptive histogram equalization (CLAHE), cropped using Otsu thresholding, and passed to an ImageNet-pretrained ResNet50 that we fine-tuned in two stages. For comparison, we trained a custom four-layer CNN, VGG16 and MobileNetV2 on the same stratified 80/10/10 split of a public dataset of 6,000 images. ResNet50 was clearly the strongest: of 600 held-out test images it classified 522 correctly (87.00% accuracy; 95% CI 84.1–89.5%), with a macro F1-score of 86.9% and a macro ROC-AUC of 0.9918, whereas the other three models reached only 48.20% to 74.10%. Performance was uneven across classes, however. Recall ranged from 95.8% for AB+ down to 73.2% for O−, and 9.0% of Rh-negative images were labelled Rh-positive. These results should also be read as within-dataset performance only, since the split was made by image rather than by person, all images came from one source, and the sensor pathway has not yet been checked against laboratory results. HemoPrint's output is therefore a prediction of blood group, not a clinical determination, and it cannot replace serological testing.
关键词
Blood group prediction, fingerprint biometrics, deep learning, transfer learning, ResNet50, convolutional neural networks, CLAHE, non-invasive diagnostics
稿件作者
ARUNESH PRABEN A S
Rajalakshmi Engineering College
Jonathan Azel Daniel
Rajalakshmi Engineering College
Priyanka D
Rajalakshmi Engineering College
发表评论