(Published at MICCAI 2024)
A large-scale multi-domain WBC detection benchmark along with morphological attributes of WBCs for prognosis of leukemia is introduced morphological attributes, recommended by hematologists for prognosis of leukemia, To facilitate future research, we have constructed extensive WBC’s detection and UDA baselines, A multi-headed WBC detection and morphological attribute prediction architecture are introduced.
To achieve explainable WBC detection, we propose to use a multi-headed WBC detector. We firstly apply recent object detectors from different domains including one-stage, YOLOv5, two-stage (Spare-Faster-RCNN) detectors, and transformer (DINO). We chose these methods because of their good detection results, efficiency, and low memory consumption. Due to the overall better result of YOLOv5 on our datasets, we have extended the YOLOv5 for attribute prediction. In YOLOv5 architecture, we added a lightweight attribute head for the prediction of WBC’s morphology. The architecture of attribute head is based on a small Convolutional Neural Network(CNN). For attribute prediction, we want to capture low-level visual details, therefore, we fuse features from two initial layers which caries structural and semantically enriched information. Specifically, we extract the different level features from YOLOv5x-backbone, extensive experiments were performed based upon different combinations of feature fusion after performing ROI align. The experimental results revealed that initial two layer features contain structural and semantically enriched information which are effective for attribute prediction.
To train the attribute head with YOLOv5 heads, asymmetric loss is employed. The YOLOv5 method-based object detections and attribute head prediction collectively predict the cell and its morphology which gives the explainable reasoning.
These predictions have been registered in a WBC morphology bank and blood film level descriptions have been generated based on the most frequently appearing WBC type (recommended by hematologists). As in the collected dataset, the blood film level data have been captured using a microscope, therefore a blood film level explainability can be produced as shown in Figure 1, which is a helpful second opinion for the hematologists
To gain a comprehensive understanding of the Leukemia prognosis and its impact, we discussed with several healthcare professionals from different working environments and finalized the WBC types and their morphological attributes. In this dataset collection, the PBFs are utilized from the diagnostic lab, ensuring patient consent, and incorporating the hematologist’s marks in the region from the monolayer area. To capture images, we utilize two distinct microscopes the high-cost (Olympus CX23) and the low-cost (XSZ-107BN) – in conjunction with two separate cameras, namely the HD1500T (HCM), ZZCAT 5MP (LCM) and the Honor 9x Lite mobile camera (HCM, LCM). It is quite challenging to locate the same patch on the PBF when employing different microscopes and resolutions. To address this inherent challenge, we initiated the capturing process by setting the field of view (FoV) at 10x, and 40x with an approximate 20% overlap, maintaining a fixed x-axis stage scale. At 100x magnification, we captured the FOV containing WBCs without any overlap, ensuring the distinct representation of individual WBCs. This process was repeated both for HCM and LCM.
The set of rules for WBC morphology varies depending on the hematologists. To enhance prognostic assistance, hematologists
identified the 14 types of WBC and considered seven key morphological attributes for a well-informed prognosis. To annotate the WBC type and morphology attributes, hematologists reviewed the subsets of the captured images and then selected the most structural and best quality images from the given subset (HD1500T paired camera at 100x using HCM) of the LeukemiaAttri dataset. For quality control, two hematologists annotate each cell with the consultation. The detail of some types of WBC with the morphological information is shown in the Figure 2 , where A) monoblasts, B) monocytes, and C) myelocytes cells exhibit mostly similar morphological characteristics as they originate from the myeloid lineage.
PhD Fellow, Intelligent Machines Lab, ITU, Lahore, Pakistan, ITU, Lahore, Pakistan
Email: phdcs23002@itu.edu.pk
LinkedIn: Abdul Rehman
BibTex
@inproceedings{rehman2024large,
title={A large-scale multi domain leukemia dataset for the white blood cells detection with morphological attributes for explainability},
author={Rehman, Abdul and Meraj, Talha and Minhas, Aiman Mahmood and Imran, Ayisha and Ali, Mohsen and Sultani, Waqas},
booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
pages={553--563},
year={2024},
organization={Springer}
}