Journal of Artificial Intelligence in Medical Applications

Accurate classification of blood cells is crucial for early diagnosis and monitoring of hematological disorders. However, manual microscopy is time-consuming, subjective, and prone to inter-observer variability. Deep learning has shown potential in automating this task, but its high computational demands limit its use in resource-constrained settings. This study aims to develop an efficient, hybrid quantum-classical model for blood cell classification using quantum simulation to reduce computational complexity without compromising accuracy. We propose a hybrid framework that integrates a 6-qubit variational quantum circuit (VQC) with a classical EfficientNet-B0 model. The model was trained and tested on the full BloodMNIST dataset, which includes 17,092 microscopic images across eight blood cell types. The proposed model achieved a test accuracy of 96.58% and a validation accuracy of 96.90%. F1-scores ranged from 91.97% to 99.84%, with notable results in neutrophils (98.80%) and lymphocytes (98.36%). These findings confirm the feasibility of combining quantum and classical approaches for medical image classification. The proposed hybrid model shows strong promise for practical use in diagnostic systems, especially in low-resource settings, and offers a pathway for future research in quantum-enhanced medical imaging.

Accurate segmentation of white blood cells (WBCs) in microscopic blood smear images is critical for diagnosing haematological disorders such as leukemia. However, manual analysis by pathologists remains labor-intensive and prone to variability, necessitating automated solutions. This study proposes a color-based WBCs segmentation framework leveraging the HSV color space and Differential Evolution (DE) optimization approach to address challenges posed by heterogeneous staining, overlapping cells, and imaging artefacts. The proposed WBCs segmentation framework operates in three stages: pre-processing via contrast enhancement and HSV transformation, region-based segmentation using DE-optimized dual-thresholding, and post-processing with morphological operations. The DE approach can learn to adaptively determine best-fit HSV thresholds for the separation of WBCs without heavy dependence on annotated datasets or deep learning architectures. Experimental analysis on the BCCD and ALL-IDB2 datasets demonstrates superior performance, with IoU scores of 99.56% and 98.78%, Dice coefficients of 97.23% and 99.53%, and F1-scores of 99.44% and 99.18%, respectively. By combining stochastic optimization with color space analysis, the presented framework yields a computationally efficient, interpretable alternative to complex deep learning models, enabling robust WBCs segmentation in resource-constrained clinical environments. The study refers to the potential of the proposed framework for incorporation into computer-automated diagnostic systems, which would minimize the dependence on manual analysis and enhance diagnostic throughput and consistency.

Medical image classification poses significant challenges due to the complex and nonlinear patterns present in imaging data. Traditional deep learning models often struggle to capture the intricate relationships inherent in such data, especially under conditions of noise and training instability. To address these limitations, this study proposes a novel Hybrid Quantum Convolutional Neural Network (HQCNN) integrated with a Gradient Variance-Controlled Optimizer (GVCO). The aim is to enhance diagnostic accuracy by combining classical convolutional layers for shallow feature extraction with a parameterized quantum circuit (PQC) that enables richer and more discriminative feature representations. The GVCO dynamically adjusts learning rates and momentum coefficients based on real-time gradient variance, improving convergence speed and generalization performance. The proposed HQCNN-GVCO model was evaluated on two public brain imaging datasets: the Kaggle Brain MRI dataset and the REMBRANDT dataset. It achieved classification accuracies of 99.2% and 98.2%, respectively. Compared to baseline HQCNN and other state-of-the-art models, HQCNN-GVCO demonstrated superior and more consistent performance. These findings highlight the potential of integrating quantum computing techniques with adaptive optimization strategies to advance reliable and accurate medical image analysis.

Gestational Diabetes Mellitus poses important health risks to both mother and child if not detected and managed early. The timely detection is important for minimizing outcomes ranging from preeclampsia and high baby birth weight to later onset of type 2 diabetes. This manuscript outlines a robust and interpretable machine learning pipeline to predict GDM, utilizing clinical data acquired from pregnant women from Kurdistan region of Iraq. The SMOTE technique is incorporated for tackling the imbalanced class problem over the dataset and to improve the model performance to detect minority class probes of GDM. We also use the Light Gradient Boosting Machine (LightGBM) as our classifier because of its fast speed and high accuracy; and the hyper-parameters are searched by automatically by Optuna, a widely used hyper-parameter optimization framework. The model was tested in two different ratios of train-test split (70%–30% and 80%–20%), and it achieved high accuracy consistently (the highest model performance accuracy reached up to 91%). These results demonstrate the advanced predictive power of the optimized model over the baseline and pragmatic application approaches. Practicality of this approach will lead to early detection of GDM timely interventions for pregnancy outcome thereby encouraging an approach against mother and fetus health.

Diabetic retinopathy (DR) is a leading cause of vision loss worldwide, and early detection is crucial to prevent irreversible blindness. This study aims to develop an effective deep learning framework for the automated multi-class classification of DR severity using color fundus images. We adopt EfficientNet-B2 as the backbone architecture, pre-trained on ImageNet and fine-tuned under various configurations to identify the optimal number of trainable layers. A customized preprocessing pipeline was introduced, incorporating median filtering, contrast enhancement using CLAHE, and circular cropping to emphasize retinal features while removing irrelevant background. To address class imbalance, aggressive data augmentation techniques were applied, resulting in a balanced dataset of 26,000 images across five DR stages. The model was trained and evaluated on a stratified 70:10:20 split of training, validation, and test sets. Among the fine-tuning strategies tested, unfreezing the last 100 layers of EfficientNet-B2 combined with the proposed preprocessing achieved the best performance, with 98.23% accuracy, 98.24% precision, 98.23% recall, and a 98.23% F1-score. These results demonstrate that targeted data preprocessing and selective fine-tuning significantly enhance classification accuracy in DR screening. The results confirm that combining tailored preprocessing with fine-tuned EfficientNet-B2 enables accurate and efficient multi-class DR classification. This framework offers strong potential for reliable, scalable deployment in clinical screening systems.

Accurately predicting the appropriate referral department from clinical records is essential for improving triage efficiency, resource allocation, and overall hospital workflow. In this study, we present an end-to-end pipeline for department prediction using unstructured Chinese electronic health records (EHRs) from the publicly available MedCD dataset. The dataset contains detailed admission notes written in natural Chinese, which often introduce challenges such as class imbalance and linguistic variability. To address these issues, we applied a series of preprocessing steps, including the removal of incomplete records, the consolidation of closely related departments (e.g., orthopedic sub-units), and the exclusion of labels with very low sample frequency. To further mitigate class imbalance, we used nlpcda, a Chinese-specific data augmentation toolkit, which generated additional samples for underrepresented classes through synonym substitution and textual perturbation. The augmented dataset was then balanced using class-weighted loss functions. For classification, we fine-tuned the Chinese-BERT-wwm-ext model on the processed dataset. Evaluation across multiple runs demonstrated clear performance gains, particularly in macro and weighted F1-scores. Overall accuracy reached 91%, while the macro F1-score improved from 0.66 (baseline) to 0.89 after augmentation. Departments with historically limited data—such as Radiology, Emergency Medicine, and Pain Management—showed marked improvement. In addition, we reported Top-3 prediction accuracy to better reflect real-world triage scenarios, where the correct department may reasonably be among the top three recommendations. These findings highlight the value of combining language-specific data augmentation with fine-tuned deep learning models for clinical text classification, especially in low-resource healthcare settings.

Accurate classification of kidney tumors from CT images remains a significant challenge due to the complexity of anatomical structures and the visual similarity between normal and pathological tissues. While conventional CNN-based methods have achieved promising results, they often suffer from overfitting and limited generalization, especially when trained on fixed 80-20 splits. This study aims to design a lightweight yet accurate framework for kidney tumor classification using a Vision Transformer Transfer Learning approach. The proposed model leverages a ViT-Tiny architecture pre-trained on ImageNet and fine-tuned on a carefully curated subset of the CT Kidney Dataset, distinguishing between normal and tumor cases. Preprocessing steps and optimization strategies were employed to enhance both model performance and computational efficiency. To improve robustness and avoid the overfitting, the model was evaluated using both 5-fold and 10-fold cross-validation, along with a more challenging 69%–31% train–test split. Experimental results show that the VTTL framework achieved a classification accuracy of 99.32%, an F1-score of 97.97%, a precision of 98.28%, and a recall of 97.97%, outperforming or matching several state-of-the-art CNN, hybrid, and transformer-based methods. The VTTL model offers a powerful and efficient solution for automated kidney tumor detection, making it highly suitable for deployment in low-resource clinical environments and online diagnostic platforms.

Diabetes has become a growing health concern worldwide, posing serious risks to people’s well-being. The condition develops when blood sugar levels remain consistently high, often as a result of factors such as physical inactivity, poor dietary habits, excess body weight, and other related influences. This study proposed a framework for detection of diabetes and non-diabetes cases based on machine learning approach and Pima Indian diabetes dataset (PID). The dataset consist of eight main indicators for detection within 764 samples. Random Forest (RF) is proposed as machine learning model for detection process. Due to avoid computational load and improve detection performance a feature selection method called Information Gain (IG) is used. Furthermore, due to misleading results challenge generated by unbalanced data of PID. This research employs SMOTE (Synthetic Minority Over-sampling Technique) to handle unbalanced data issue. The result of this study indicates that detection process is more effective based on feature selection method and balanced dataset. To mention, RF model scored 84% for accuracy, precision, recall, and F-score. Beside no over-fitting or underfitting issue is observed.

The increasing use of digital brain imaging in clinical diagnosis and telemedicine has raised serious concerns regarding the security and privacy of sensitive patient data. Traditional cryptographic methods may fail to provide adequate protection when handling large image sets or when additional semantic obfuscation is required. In this paper, we present a lightweight cryptographic framework for securing medical brain MRI images, based on a hybrid bio-inspired and cryptographic methodology. The architecture combines an RNA-based encoding scheme and a Substitution Permutation Network (SPN), which with high probability ensures high confusion and diffusion in the ciphertext and is compatible with biomedical data formats. Brain MRI images are mapped to a symbolically encoded RNA representation, placing each 8-bit grayscale pixel into a six-character RNA block, via custom encoding tables. The result of this representation is encrypted by a multi-round SPN cipher using a biologically inspired S-box, and a RNA-specific permutation plan. The decryption process merely inverts the SPN operations and the original image reinvents itself with no loss. Experimentally, the proposed method was tested on a Kaggle data set of brain tumor MRI images with glioma, meningioma, pituitary, and no-tumor cases. Findings indicate that 256×256 image encryption and decryption can be achieved in less than one second, which makes the framework appropriate to the telemedicine application of encryption and decryption in real time. In general, the encryption process proposed can be considered an effective and very secure system of protection of sensitive medical images.

Measure of brain dimensions is an important metric for discovering brain abnormalities associated with certain diseases. Using an animal model to measure brain dimensions, followed by a discovery of possible genes responsible for changes in the dimensions, enables better understanding of the underlying genetic cause of the brain’s abnormalities in humans. This paper presents a three-stage automatic solution for estimating brain dimensions from dorsal-ventral X-ray images of mouse heads based on machine learning techniques. At the first stage, a skull modelling technique based on circular Hough transform is used to determine region of interests of the key landmarks on the skull. In stage 2, the procedures for the width and length landmarks ROIs estimation based on a circle model are presented. The landmark locations are then obtained in stage 3 using the random forests classification method based on texture features derived from the local binary patterns and histogram of oriented gradients and then brain dimensions are measured in terms of relevant width and length of the skull. Experimental results on a set of X-ray head images of 798 wild type and 6,090 mutant mice have demonstrated the effectiveness of the proposed solution in separating images of abnormal brain dimensions from those of normal ones. Furthermore, the estimated dimensions have also led to identifying 7 possible genes responsible for changes in the brain size of the mutant mice, some of which were already linked to Down’s syndromes, which further demonstrates the practical usefulness of the proposed solution.