Moderate evidence
More
AI reading 500 cell images per patient sorts leukemia subtypes with 88% accuracy in a test
Study title: Person-Based Progression in Acute Myeloid Leukemia Classification: A Multi-View Convolutional Neural Network Approach
- Evidence type
- Benchmark or lab evaluation
- Sample
- n = 81,214
- Sources
- 1 sourcePeer-reviewed
- Published
- 1 min read
- Truvace record
- TRV-2026-1264cite or check it
Study title: Person-Based Progression in Acute Myeloid Leukemia Classification: A Multi-View Convolutional Neural Network Approach
By the publication date of 2026-10-01, researchers had developed and evaluated a person-based multi-view CNN approach for AML subtype classification using 81,214 single-cell images from 189 individuals. The 500-view MV-CNN achieved accuracy 0.8783, F1 0.8622, sensitivity 0.8637, specificity 0.8774, precision 0.8616 and MCC 0.8441, outperforming a single-view CNN baseline in tenfold cross-validation.
Read the study abstract
Acute myeloid leukemia (AML) is a hematologic malignancy in which accurate subtype identification is important for treatment planning and prognosis. This study developed a deep learning approach for AML subtype classification using multiple single-cell images from each patient rather than conventional instance-based classification. An open-source dataset containing 81,214 single-cell images from 189 individuals was analyzed. The dataset included four genetically defined AML subtypes and a healthy control group.…

The gain
In tenfold cross-validation on 81,214 cell images from 189 people, the 500-view model reached 0.8783 accuracy and 0.8622 F1, beating a single-image CNN baseline that scored lowest.
In brief
The result matters because accurate subtype identification is important for treatment planning and prognosis in AML, and patient-level aggregation of multiple cell images improved classification over conventional instance-based analysis. What remains uncertain is clinical translation, as the source describes performance on an open-source dataset and supports potential for hematologic image classification rather than reporting prospective clinical deployment or patient outcomes.
Main points
- Study used open-source dataset of 81,214 single-cell images from 189 individuals covering four genetically defined AML subtypes and healthy controls.
- Two MV-CNN models were tested: 99-view model based on random sampling and 500-view model using image augmentation when fewer than 500 images were available per patient.
- Performance assessed by tenfold cross-validation on accuracy, F1, sensitivity, specificity, precision, and MCC; single-view CNN yielded the lowest values.
The full story
The comparison was a single-view CNN versus two multi-view approaches, with the 99-view model showing lower but comparable performance to the 500-view model.
Dataset included four genetically defined AML subtypes plus a healthy control group, and evaluation used tenfold cross-validation across six metrics including Matthews correlation coefficient.
Sources
1 on file- Journal of Imaging Informatics in Medicinedoi:10.1007/s10278-026-02371-7
Talk it over
The 500-view model reached 0.8783 accuracy and beat the single-view CNN across the metrics. What does that tell you about using many cell images per person rather than one?
AI reading 500 cell images per patient sorts leukemia subtypes with 88% accuracy in a testhttps://truvace.com/item/person-based-progression-in-acute-myeloid-leukemia-classification-a-multi-view-c
The 500-view model reached 0.8783 accuracy and beat the single-view CNN across the metrics. What does that tell you about using many cell images per person rather than one?
Answer in the threadThe data came from an open-source dataset of 81,214 images from 189 individuals. Who or what might be missing from that sample, and how could it change the results in a clinic?
Answer in the threadThe source reports no prospective clinical deployment or patient outcomes. What would you want to see from a real hospital trial before relying on this for AML treatment planning?
Answer in the thread
If you have more time
The 99-view model performed lower but comparable to the 500-view model. What would you look up to judge whether the extra views are worth the added effort?
Answer in the threadIf you were a patient with suspected AML, how would you want a doctor to describe the role of this kind of image classification in your care?
Answer in the thread
ace
What readers say