Uni-Hema

Uni-Hema is a unified, multi-task, multi-modal model for digital hematopathology, the microscopic analysis of blood and bone marrow cells used to diagnose blood disorders. Where most computational tools in this space are trained for a single task on a single dataset, Uni-Hema is designed to perform cell detection, classification, segmentation, morphology prediction, masked language modeling, and visual question answering (VQA) within one framework, spanning malignant disorders (leukemia), infectious conditions (malaria), and non-malignant red blood cell disorders such as sickle cell disease and anemia.

The model was introduced in a November 2025 preprint by Abdul Rehman, Iqra Rasool, Ayisha Imran, Mohsen Ali, and Waqas Sultani, a collaboration between the Information Technology University of the Punjab (Lahore) and Chughtai Lab (Lahore). Its central contribution is bridging the gap between single-cell analysis and clinical reasoning: rather than only labeling cells, Uni-Hema couples pixel-level perception with text-based outputs so that predictions are accompanied by interpretable, morphologically grounded descriptions.

By consolidating tasks that are usually handled by separate specialist models, Uni-Hema aims to provide a single backbone that generalizes across hematological diseases and data modalities, an approach well-suited to resource-constrained diagnostic settings where maintaining many task-specific pipelines is impractical.

#Key Features

• Multi-task unified architecture: A single model handles detection, classification, segmentation, morphology prediction, masked language modeling, and VQA, removing the need to train and deploy separate specialist networks per task.
• Hema-Former multimodal module: A dedicated module bridges visual and textual representations at multiple hierarchy levels using four learnable sub-modules: cross-modal fusion, text-guided visual refinement, a single-cell feature extractor, and a query-guided mask former.
• Cross-disease coverage: Trained across leukemia, malaria, sickle cell disease, and anemia, the model spans malignant, infectious, and non-malignant red blood cell disorders within one framework.
• Interpretable single-cell reasoning: Outputs include morphologically relevant textual descriptions, supporting interpretation at the single-cell level rather than opaque labels alone.
• Broad public-data pretraining: Built on 46 publicly available hematology datasets totaling more than 700,000 images, improving generalization to unseen benchmarks.

#Technical Details

Uni-Hema is composed of six principal modules: a ResNet-50 CNN image backbone for multi-scale spatial features, a six-layer transformer image encoder, a T5-based text encoder, a DINO-DETR image decoder for detection and morphology tasks, an autoregressive transformer text decoder, and the Hema-Former module that fuses the visual and textual streams. Training draws on roughly 700K images across 46 datasets, split approximately into 17 classification datasets (~320K images), 18 detection datasets (~85K images), and 11 segmentation datasets (~221K images), plus about 22K question-answer pairs and 7K masked language modeling samples. Reported results include mAP50 of 49.8 for leukemia cell detection (H_100x_C2), F1 of 98.8 for single-cell classification on Raabin, a Dice score of 99.9 for segmentation on the Elsafty anemia set, F1 of 83.6 for field-of-view morphology, BLEU-4 of 56.4 on WBCAtt-VQA, and BLEU-4 of 79.8 on masked language modeling, with a mean classification F1 of 90.8 on unseen datasets. The authors report comparable or superior performance to single-task, single-dataset baselines.

#Applications

Uni-Hema targets diagnostic hematology workflows where a pathologist must detect, count, classify, and characterize cells across smears for diseases such as leukemia, malaria, and sickle cell disease. Because it consolidates many tasks into one model and produces interpretable, morphology-aware text, it is well-suited to decision support, second-read screening, and education, and is especially relevant to resource-limited clinical settings that cannot maintain numerous task-specific systems. Researchers building hematopathology benchmarks or downstream tools may also use it as a multimodal foundation backbone.

#Impact

Uni-Hema represents an effort to move digital hematopathology from fragmented, single-task models toward unified multimodal foundation models that reason across diseases and tasks, mirroring trends seen in computational pathology and vision-language modeling more broadly. Its emphasis on interpretable single-cell outputs and on disease coverage relevant to low-resource regions is a notable contribution. Important caveats apply: the work is a preprint and has not yet been peer reviewed, and as of this writing the authors state that code "will be made publicly available" but no repository, released weights, or license have been published, so reproducibility and independent evaluation remain pending.