CoTCoNet: An Optimized Coupled Transformer-Convolutional Network with an Adaptive Graph Reconstruction for Leukemia Detection

TL;DR

CoTCoNet represents a significant advancement in leukemia classification, offering enhanced accuracy and efficiency compared to traditional methods, and provides deeper insights into its decision-making process, further solidifying its potential in clinical settings.

Abstract

Swift and accurate blood smear analysis is an effective diagnostic method for leukemia and other hematological malignancies. However, manual leukocyte count and morphological evaluation using a microscope is time-consuming and prone to errors. Conventional image processing methods also exhibit limitations in differentiating cells due to the visual similarity between malignant and benign cell morphology. This limitation is further compounded by the skewed training data that hinders the extraction of reliable and pertinent features. In response to these challenges, we propose an optimized Coupled Transformer Convolutional Network (CoTCoNet) framework for the classification of leukemia, which employs a well-designed transformer integrated with a deep convolutional network to effectively capture comprehensive global features and scalable spatial patterns, enabling the identification of complex and large-scale hematological features. Further, the framework incorporates a graph-based feature reconstruction module to reveal the hidden or unobserved hard-to-see biological features of leukocyte cells and employs a Population-based Meta-Heuristic Algorithm for feature selection and optimization. To mitigate data imbalance issues, we employ a synthetic leukocyte generator. In the evaluation phase, we initially assess CoTCoNet on a dataset containing 16,982 annotated cells, and it achieves remarkable accuracy and F1-Score rates of 0.9894 and 0.9893, respectively. To broaden the generalizability of our model, we evaluate it across four publicly available diverse datasets, which include the aforementioned dataset. This evaluation demonstrates that our method outperforms current state-of-the-art approaches. We also incorporate an explainability approach in the form of feature visualization closely aligned with cell annotations to provide a deeper understanding of the framework.

Figures (11)

• Figure 1: Illustration of sample images depicting the presence or absence of leukemic cells. The first row denotes the Whole Slide Image (WSI) of the blood samples, while the last row shows a zoomed-in view of WSI containing leukocytes. This figure highlights the indistinguishability of cell images across classes.
• Figure 2: Proposed Coupled Transformer Convolution Network (CoTCoNet) for Leukemia Detection. Firstly, the leukocyte cell images are inputted into (A) for pre-processing and segmentation. In step (A), the fusion of CLAHE and Sharpen image enhancement technique is applied, followed by Region of Interest Identification. Then, images are segmented by utilizing LeuSAM. The output of Step (A) is fed to the proposed GAN-driven cell synthesis step (B). The outcome of step (B) comprises synthetic cell images produced by the GAN adhering to the real images. In step (C), CoTCoNet first isolates the global and spatial features and then proceeds to regenerate them through graph-based feature reconstruction techniques. At step (D), a meta-heuristic sine-cosine algorithm optimizes the feature. Finally, step (E) classifies the optimized features into normal or leukemia. The effectiveness of the proposed framework is assessed on four different datasets.
• Figure 3: Proposed Generative Adversarial Network-driven architecture (LeuGAN) for synthetic leukocyte cell generation. Our approach involves a $128$-dimensional noise vector $z$, Generator $\mathcal{G}$ and corresponding Discriminator $\mathcal{D}$. Specifically, $\mathcal{G}$ uses the noise vector to craft synthetic leukocyte blood cells, while $\mathcal{D}$ evaluates both real and generated leukocytes combined. The network employs the generator-adversarial loss function $\mathcal{L}_{GAN}(\mathcal{D,G})$ to maintain consistency effectively.
• Figure 4: Architecture of the proposed Global Feature Module (top) and details of the Transformer Layer block (bottom). The input image is first split into equal-sized patches and flattened. Then, each patch is projected into a feature space where a transformer layer block processes them to extract global features.
• Figure 5: Architecture of the proposed Spatial Feature Module, comprising five Spatial Feature Learning (SFL) blocks, represented by $\varphi_{S\!F\!L(i)}$, $i \in \left[1,2,3,4,5 \right]$. The numerical values at the bottom indicate the number of filters employed in each SFL block.