Attention guided segmentation and multi resolution transformer learning for blood cancer detection

Abstract Purpose Blood cell cancers, including leukemia, lymphoma, and multiple myeloma, disrupt the hematopoietic system, necessitating precise and early diagnosis for effective treatment. Traditional diagnostic methods face limitations in sensitivity, specificity, and real-time applicability. Approach This study presents an integrated framework combining attention-guided segmentation, multi-resolution transformer-based feature extraction, and attention-enhanced classification for blood-cell image analysis. Using a publicly sourced dataset of 500 class-labelled blood-smear images, the framework incorporates UNet++ with Attention Gates for segmentation, multi-resolution DeiT for feature extraction, and AE-MLP for AML, ALL, and healthy-cell classification. Results The final attention-guided UNet++ configuration achieved a Dice coefficient of 0.95, an IoU of 0.88, precision of 0.96, recall of 0.93, and pixel-level segmentation accuracy of 0.97. Multi-resolution DeiT representation learning was associated with improved boundary delineation relative to the evaluated segmentation comparators and achieved a mean Grad-CAM-based Attention Focus Score of 92.4%, indicating that 92.4% of the normalized class-associated activation was concentrated within the designated cellular region. The AE-MLP classification stage achieved an overall accuracy of 96.0%, a rare blast-cell sensitivity of 94.7%, and an AUC-ROC of 0.95. The evaluated pipeline required 1.2 s per image during single-image inference and achieved a batched throughput of 80 images per minute on the portable-GPU configuration. Error-guided fine-tuning reduced the observed misclassification rate from 12 to 7%, corresponding to an absolute reduction of 5 percentage points. The framework achieved higher performance than the evaluated CNN-based and vanilla Vision Transformer baselines under the common experimental protocol; however, external clinical validation remains necessary. Conclusion The findings demonstrate the technical potential of integrating attention-guided segmentation, transformer-based representation learning, and interpretable classification for computer-assisted analysis of AML, ALL, and healthy blood-cell images.

Authors

Institutions

Publication Details

Journal
Discover Artificial Intelligence
Published
2026-11-17
DOI
https://doi.org/10.1007/s44163-026-02492-6
Primary Topic
Digital Imaging for Blood Diseases
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Attention guided segmentation and multi resolution transformer learning for blood cancer detection

Vinay Kukreja, Vivek Bhardwaj, Ayush Dogra, Eshika Jain et al.
7 links
Discover Artificial Intelligence
Digital Imaging for Blood Diseases
article

Attention guided segmentation and multi resolution transformer learning for blood cancer detection

Vinay Kukreja, Vivek Bhardwaj, Ayush Dogra, Eshika Jain, Pratham Kaushik, Ajay Kumar
article en

Abstract

Abstract Purpose Blood cell cancers, including leukemia, lymphoma, and multiple myeloma, disrupt the hematopoietic system, necessitating precise and early diagnosis for effective treatment. Traditional diagnostic methods face limitations in sensitivity, specificity, and real-time applicability. Approach This study presents an integrated framework combining attention-guided segmentation, multi-resolution transformer-based feature extraction, and attention-enhanced classification for blood-cell image analysis. Using a publicly sourced dataset of 500 class-labelled blood-smear images, the framework incorporates UNet++ with Attention Gates for segmentation, multi-resolution DeiT for feature extraction, and AE-MLP for AML, ALL, and healthy-cell classification. Results The final attention-guided UNet++ configuration achieved a Dice coefficient of 0.95, an IoU of 0.88, precision of 0.96, recall of 0.93, and pixel-level segmentation accuracy of 0.97. Multi-resolution DeiT representation learning was associated with improved boundary delineation relative to the evaluated segmentation comparators and achieved a mean Grad-CAM-based Attention Focus Score of 92.4%, indicating that 92.4% of the normalized class-associated activation was concentrated within the designated cellular region. The AE-MLP classification stage achieved an overall accuracy of 96.0%, a rare blast-cell sensitivity of 94.7%, and an AUC-ROC of 0.95. The evaluated pipeline required 1.2 s per image during single-image inference and achieved a batched throughput of 80 images per minute on the portable-GPU configuration. Error-guided fine-tuning reduced the observed misclassification rate from 12 to 7%, corresponding to an absolute reduction of 5 percentage points. The framework achieved higher performance than the evaluated CNN-based and vanilla Vision Transformer baselines under the common experimental protocol; however, external clinical validation remains necessary. Conclusion The findings demonstrate the technical potential of integrating attention-guided segmentation, transformer-based representation learning, and interpretable classification for computer-assisted analysis of AML, ALL, and healthy blood-cell images.

Discover Artificial IntelligenceVol. 6(1)
Amity University (IN), Amity University (AE), Manipal University Jaipur (IN), Chitkara University (IN)
Good health and well-being
Digital Imaging for Blood Diseases
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon