Research Hub · CNN ABO Blood Group Classification

CNN‑Based Blood Group Classification

Automated ABO blood group detection from microscopic / agglutination images using deep learning. This project systematically addresses the lack of public ABO-labeled datasets, introduces explainable AI (Grad‑CAM), and benchmarks modern lightweight architectures. Built on a 30‑day structured research plan with IEEE‑format publication output.

📊 4/8 classes (ABO / Rh+) 🧠 MobileNetV2 + EfficientNet 🔍 Grad‑CAM explainability 📄 IEEE template ready

Available Public Datasets

🏥 PBC Dataset (Barcelona) — 17,092 cells, 8 types, 360×363px.
DOI:10.1016/j.dib.2020.105474
✅ Primary recommendation for morphology baseline.

🩸 BCCD — 364 annotated images (WBC,RBC,platelets). Object detection style.

📦 TXL-PBC (Nature 2025) — Merged 4 sources, 1,260 images, 18k annotations.

🔬 MLL Dataset (Nature 2025) — >40,000 single-cell images, 18 classes, largest hematology lab.

🧬 Kaggle: Blood Cells Image Dataset — no direct ABO labels, usable for transfer learning.

📄 IEEE 2022 Custom ABO Dataset — self-created, unpublished, 96.7% accuracy (most relevant prior work).

📌 Dataset strategy for your paper: Since no large public ABO-labeled microscopic dataset exists, we can either (a) construct a custom agglutination card dataset, or (b) frame the pipeline using PBC dataset for morphology + ABO inference as a novel extension. The timeline incorporates custom dataset creation and preprocessing.

Prior Work & Research Gaps

Key directly related paper: IEEE ICCCNT 2022 — custom CNN on ABO images, 96.7% accuracy, but no XAI, no Rh factor, no modern backbones.

Gap 1 — No public ABO-labeled microscopic dataset → your contribution can be a reproducible dataset.

Gap 2 — No explainability in ABO classification → add Grad‑CAM, SHAP, LIME.

Gap 3 — No comparison of MobileNetV2 / EfficientNet for ABO task.

Gap 4 — Rh factor (8-class: A+,A-,B+,B-,AB+,AB-,O+,O-) unexplored.

Gap 5 — Class imbalance (O common, AB rare) not addressed.

Gap 6 — Geographic/population bias, lack of generalisation discussion.

Accuracy Benchmarks to Beat

Paper / Model	Task	Accuracy
IEEE 2022 (Custom CNN)	ABO blood group	96.7%
Islam et al. 2024 (Optimized CNN+XAI)	WBC classification	99.12%
Gavas & Olpadkar 2021 (ensemble)	PBC cell type	99.51%
MobileNetV2 (2024 benchmark)	WBC	~98.58%
EfficientNet-B3 (2025)	Leukemia ALL	94.30% F1

🎯 Your target: surpass 96.7% using improved preprocessing (CLAHE), transfer learning (MobileNetV2/EfficientNetB0), Grad‑CAM explainability, and proper validation.

Positioning Statement (Abstract/Intro)

“While prior studies have demonstrated CNN‑based classification of peripheral blood cell subtypes with high accuracy, the specific problem of automated ABO blood group classification from microscopic images remains underexplored, with limited publicly available labeled datasets and no work addressing explainability or comparison of modern lightweight architectures. This study proposes a CNN‑based pipeline for ABO blood group classification incorporating optimized preprocessing, transfer learning via MobileNetV2/EfficientNet, and Grad‑CAM‑based explainability, providing a transparent, efficient, and reproducible framework for AI‑assisted hematological diagnostics.”

Suggested Paper Structure (IEEE Format)

1. Abstract (~250 words)
2. Introduction – importance of blood typing, AI motivation
3. Related Work – 10–15 citations (PBC, IEEE 2022, XAI papers)
4. Dataset Description – custom / PBC / agglutination, class distribution
5. Proposed Methodology – preprocessing, CNN + transfer learning, Grad‑CAM
6. Experimental Results – accuracy, precision, recall, F1, confusion matrix
7. Comparative Analysis – vs AlexNet, LeNet5, MobileNetV2 baseline
8. Discussion – gaps, limitations, interpretability
9. Conclusion & Future Work – Rh factor, real-time deployment
10. References (IEEE format, 15–20 entries)

📚 Recommended citations: Acevedo 2020, IEEE 2022 ABO, Islam 2024, Gavas 2021, PLOS ONE scoping review, etc.