Education Academy Logo
Journal Logo

Conference Abstracts - 2026 National Summit on Hematologic Cancers

Vol. 6, Issue Supplement 2, 2026 · S1-2

Automatic Detection of Acute Lymphoblastic Leukemia Using ALNET Deep Learning Model

Doreen Ninsiima, BME,Joshua Biryomumeisho, MSc,Gabriel Irumba Rumwaya, BME,Johnes Obungoloch, PhD,William Wasswa, PhD

Acute Lymphoblastic LeukemiaSmartphone MicroscopyPediatric OncologyArtificial IntelligenceDeep LearningDigital Health

Submission received: 2026-08-14 / Accepted: 2026-08-25 / Published: 2026-09-28

CCBY-SA-4.0
Publication: IJCCDhttps://doi.org/10.53876/001a.129824
2

Abstract

Background

Acute lymphoblastic leukemia (ALL) accounts for 75-80% of childhood leukemias, with low- and middle-income countries (LMICs) bearing the greatest burden. Uganda reports 2.67 leukemia deaths per 100,000 and a devastating 16.5% 1-year survival at the Uganda Cancer Institute. Diagnostic delays stem from scarce hematopathologists (<1 per 100,000), manual microscopy, and symptom overlap with malaria. No AI-based ALL diagnostic tool has been validated on African clinical samples. Hypothesis: A CNN-based smartphone-microscopy diagnostic tool validated on Ugandan clinical samples can achieve ≥90% accuracy, reducing diagnostic time from hours to minutes.

Methods

Using Design Science Research, we developed LeukNet, a CNN (ALNet architecture) trained on 260 ALL-IDB1 images (blast versus non-blast). Images were preprocessed (224×224px, normalized, augmented) with a 70-30 train-validation split. External validation used 30 blood smear slides (20 leukemic, 10 non-leukemic) from Mbarara Regional Referral Hospital, Uganda. Images were captured via compound microscope using two smartphones: Techno Camon 20 Premier (108MP) and Itel S23 (50MP), yielding 300 images/device. Two pathologists provided ground-truth labels. Performance metrics: accuracy, sensitivity, specificity, F1-score.

Results

Internal validation achieved 98% accuracy, 95.8% sensitivity, 97.1% specificity, and 96.3% F1-score. The confusion matrix showed 87% overall accuracy (blast precision: 0.96; non-blast recall: 0.97). External validation demonstrated 100% accuracy for positive smears on Techno and 90% on Itel; negative smears achieved 90% (Techno) and 80% (Itel). Errors correlated with lower resolution and reactive cell confusion. Overall external accuracy: 90%. The web platform with Mistral LLM summaries and QR verification proved operationally feasible.

Conclusion

LeukNet is the first AI-based ALL diagnostic tool validated on Ugandan clinical samples, achieving 90% external accuracy using widely available smartphones and standard microscopes. By reducing diagnostic time from hours to minutes without specialized equipment, LeukNet offers a scalable, cost-effective solution for LMICs, supporting a multi-site pilot to improve childhood ALL survival in Uganda and sub-Saharan Africa.