FreqMIL: Frequency-Aware Mamba Multi-Instance Learning for Acute Leukemia Typing

Le résumé fourni par la source

Acute leukemia (AL) is a heterogeneous hematologic malignancy, where accurate diagnosis requires large-scale data, and the weak labeling problem makes AL typing particularly challenging. To address the above issues, we propose a novel framework integrating optical time-stretch imaging flow cytometry (OTS-IFC) with frequency-aware Mamba multiple instance learning (FreqMIL). We recruit 33 clinical patients, including 7 acute lymphoblastic leukemia (ALL), 18 acute myelogenous leukemia (AML), and 8 healthy donors, and acquire 229,835 high-throughput cell images in a label-free manner. To capture heterogeneous features from low-abundance cell populations, we propose FreqMIL, which enhances feature representation in the frequency domain and models sequential dependencies, enabling robust sample-level AL typing. Experimental results demonstrate that FreqMIL outperforms six representative MIL methods and achieves 92.08% accuracy in AL typing, highlighting its potential for reliable and efficient clinical diagnosis.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.