Loading...

Media is loading
 

Degree

Bachelor of Science (Computer Science)

Department

Department of Computer Science

School

School of Mathematics and Computer Science (SMCS)

Advisor

Dr. Rizwan Ahmed Khan, Professor, Department of Computer Science

Co-Advisor

Dr. Muhammad Rizwan Alam (Industry Advisor)

Keywords

Cellular Morphological Explainability, Domain Adaptation, Annotated Hematology Dataset Curation

Abstract

Automated white blood cell classification and morphological analysis in peripheral blood smears sits at the intersection of two critical unmet needs in modern healthcare: the inaccessibility of conventional hematology analyzers in resource-constrained clinical environments due to their prohibitive cost, and the well-documented failure of deep learning models to generalize across clinical sites due to variations in staining protocols and imaging hardware , a phenomenon known as covariate shift. This project directly addresses both challenges through a comprehensive pipeline spanning data curation, morphometric analysis, domain adaptation, and vision-language modelling. First, in collaboration with Quaid-i-Azam University, Islamabad, we curated a novel, high-quality peripheral blood smear dataset from local Pakistani clinical sources, featuring precise cell-wall and nucleus polygon annotations, segmentation masks, and morphological diversity across multiple stimulation conditions, filling a significant gap in regional representation in public hematology datasets. Second, we developed an automated cell annotation and morphometric framework that computes instance-wise Nucleus-to-Cytoplasm ratios from annotated microscopy images, reducing what previously demanded months of expert manual effort to a matter of hours. Third, we propose SER-DA-MMD, a parameter-efficient unsupervised domain adaptation framework that inserts a lightweight Squeeze-Excitation Residual adapter into a frozen backbone guided by Maximum Mean Discrepancy alignment, achieving substantial accuracy gains over direct transfer with no target domain labels and no backbone retraining required. Finally, we fine-tuned a Vision-Language Model on cell morphology captions to enable natural language understanding of leukocyte structure, with a clinician-facing desktop application integrating all components currently under active development for point-of-care deployment.

Tools and Technologies Used

PyTorch, timm, Hugging Face Transformers, Gemma Vision-Language Model, scikit-image, tifffile, GradCAM, Squeeze-and-Excitation Networks, Maximum Mean Discrepancy with multi-kernel RBF kernels, Adaptive Batch Normalization, ResNet50, VGG16-BN, InceptionV3, Vision Transformer ViT-Base, AdamW optimiser, NVIDIA Tesla T4 GPU, Kaggle

Methodology

Our methodology follows a clear four‑stage pipeline designed for real clinical impact. First, we curated a novel peripheral blood smear dataset from local Pakistani sources in collaboration with Quaid‑i‑Azam University, Islamabad. We collected microscopy images across three separate experiments under both Control and stimulated conditions using PMA and FMLP at varying concentrations. Expert hematologists meticulously annotated each image by drawing polygon boundaries around cell walls and nuclei, providing perfect ground‑truth for training and evaluation. Second, building on these annotations, we built an automated cell annotation and NC‑ratio computation framework. What previously demanded months of manual effort is now completed in few days with the tool, or in a few hours when assisted by our AI model. Third, to overcome the devastating performance drop caused by covariate shift across different clinical sites, we developed SER‑DA‑MMD, a parameter‑efficient unsupervised domain adaptation framework. We keep the entire pre‑trained backbone frozen and insert a lightweight trainable Squeeze‑Excitation Residual adapter just before the classifier head. This adapter uses channel‑wise attention to suppress stain‑sensitive artefacts while amplifying morphologically informative structures. Training is guided by a hybrid loss that combines cross‑entropy on labelled source data with Maximum Mean Discrepancy (MMD) alignment on unlabelled target data. For CNN architectures, we additionally apply Adaptive Batch Normalisation (AdaBN) as a post‑training step. Fourth, we fine‑tuned a Florence‑style vision‑language model on haematology‑specific image‑caption pairs, enabling the system to generate natural language descriptions of cell morphology. A desktop application integrating all modules is currently under active development for direct clinical deployment

Document Type

Restricted Access

Submission Type

BSCS Final Year Project

Creative Commons License

Creative Commons Attribution-Noncommercial 3.0 License
This work is licensed under a Creative Commons Attribution-Noncommercial 3.0 License

Download

Share

COinS