publications

2023

1. NeurIPS NewInML 2023

   Scalable Multi-Teacher Gated Knowledge Distillation for Accurate & Efficient Leukemia Classification

   Tahsen Islam Sajon

   In NeurIPS 2023 NewInML Workshop (Non-Archival), Dec 2023

   Abs PDF Code

   Recent advancements in deep learning have driven significant progress in medical imaging, notably in automated leukemia classification. However, deploying state-of-the-art models in real-world applications can be prohibitive due to their computational and memory overhead. In this context, this study introduces a multi-teacher knowledge distillation framework, complemented by a learned gating mechanism. This gate is designed to dynamically allocate weights to the teacher models during the distillation process, promoting accurate predictions in the student model. Swin Transformer and ConvNext models serve as the chosen teachers, with the intention to harness both the robust representational capabilities of transformers and the locality-preserving attributes of convolutions. Employing this strategy to guide a lightweight MobileNetV2 student model, an evident improvement in performance was achieved over conventional training techniques

2. BIM 2023

   Attention Mechanism-Enhanced Deep CNN Architecture for Precise Multi-class Leukemia Classification

   Tahsen Islam Sajon, Barsha Roy, Md. Farukuzzaman Faruk, and 5 more authors

   In Proceedings of the 2nd International Conference on Big Data, IoT and Machine Learning, Sep 2023

   Abs PDF

   Leukemia is a life-threatening condition affecting people globally, making accurate diagnosis crucial for timely intervention. Consequently, researchers have been exploring automated methods to enable prompt action. The classification of leukemia into multiple subtypes according to WHO standards presents a unique challenge. Unlike binary classification, interclass features are highly similar, leading to misclassification. Ergo, we employ attention mechanisms to tackle this problem. Our proposed deep learning architecture combines transfer learning with attention mechanisms to classify subtypes of leukemia accurately. Using a publicly available dataset of blood cell images that adhered to WHO standards, we illustrate the potency of our approach. Our DenseNet201 with CBAM model achieves a remarkable 99.85% overall accuracy without resorting to data augmentation, surpassing previous methods on this dataset and attaining state-of-the-art results compared to other leukemia literature. To interpret the model’s decision-making process and evaluate the efficacy of the attention mechanism in identifying discriminating features, we showcase GradCAM images and intermediate layer feature maps generated from our custom CNN. The proposed approach enhances leukemia classification accuracy and efficiency, providing clinical decision-making insights.

3. ECCE 2023

   Recognition of Leukemia Sub-types Using Transfer Learning and Extraction of Distinguishable Features Using an Effective Machine Learning Approach

   Tahsen Islam Sajon, Maria Chowdhury, Azmain Yakin Srizon, and 4 more authors

   In 2023 International Conference on Electrical, Computer and Communication Engineering (ECCE), Feb 2023

   Abs PDF

   Although extensive research has been conducted on leukemia, the disease still accounts for more than 350,000 fatalities annually. Automated Leukemia diagnosis may alter the situation because actions can be taken immediately; as a result, accurate detection of Leukemia has been a subject of interest for researchers. As statistics grow and expand, the need for precise leukemia identification continues to increase. In this study, we investigated a dataset of leukemia that used the WHO classification scheme. We developed a modified DenseNet201 design that achieved an overall accuracy of 99.69% without relying on data augmentation. Additionally, we identified and validated key features for leukemia classification by utilizing three feature extraction approaches (i.e., hu moments, haralick texture and parameter-free threshold adjacency statistics) and several machine learning classifiers (i.e., Gaussian Process, Support Vector Machine, K-Nearest Neighbor or KNN, Extra Trees Classifier, and Logistic regression) that outperformed earlier feature extraction-based techniques.

4. BIM 2023

   Large Ensemble of Transfer-Learned Models for Plant Disease Recognition from Diverse Leaf Images

   Maria Chowdhury, Azmain Yakin Srizon, Tahsen Islam Sajon, and 6 more authors

   In Proceedings of the 2nd International Conference on Big Data, IoT and Machine Learning (In Press), 2023

   Abs PDF

   Plant diseases caused by microorganisms like fungi, bacteria, and insects can have devastating effects on crop productivity and quality. Detecting and identifying plant diseases is essential for effective management and sustainable agriculture. Computer vision-based image analysis employing deep learning techniques has emerged as a promising solution for efficient and timely disease identification in crops. This study aimed to enhance the classification performance of plant leaf disease images from the PlantDoc dataset by employing an ensemble CNN technique. Seven pre-trained CNN models were trained on the dataset, and a SoftMax Averaging ensemble technique was used to achieve superior plant disease classification performance. The proposed approach outperformed previous studies by a great margin, achieving an accuracy of 98.66%. The SoftMax Averaging ensemble technique allowed for a more diverse range of feature extraction and improved classification accuracy which can effectively address the challenge of plant disease classification.