Karima Bahmane

Dr. Karima Bahmane


Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
14 November 2025 in Medicina
Verified
Journal Article
Open access
PDF file
KBKarima Bahmane
Karima Bahmane
SBSambit Bhattacharya
Sambit Bhattacharya
BCBrahim Elkhalil Chaouki
Brahim Elkhalil Chaouki
Background and Objectives: Stratifying thyroid nodules according to malignancy risk is a crucial step in early diagnosis and patient care. Recently, deep learning techniques have emerged as powerful tools for medical diagnostics, particularly with convolutional neural networks (CNNs) applied to medical image classification. This study aimed to develop a new hybrid CNN model for classifying thyroid nodules using the TN5000 ultrasound image dataset. Materials and Methods: The TN5000 dataset includes 5000 ultrasound images, with 3572 malignant and 1428 benign nodules. To address the issue of class imbalance, the researchers applied an R-based anomaly data augmentation method and a GAN-based technique (G-RAN) to generate synthetic benign images, resulting in a balanced dataset for training. The model architecture was built on a pre-trained EfficientNet-B3 backbone, further enhanced with squeeze-and-excitation (SE) blocks and residual refinement modules to improve feature extraction. The task was to classify malignant nodules (labeled 1) and benign nodules (labeled 0). Results: The proposed hybrid CNN achieved strong performance, with an accuracy of 89.73%, sensitivity of 90.01%, precision of 88.23%, and an F1-score of 88.85%. The total training time was 42 min. Conclusions: The findings demonstrate that the proposed hybrid CNN model is a promising tool for thyroid nodule classification on ultrasound images. Its high diagnostic accuracy suggests that it could serve as a reliable decision-support system for clinicians, improving consistency in diagnosis and reducing human error. Future work will focus on clinical validation, explainability of the model’s decision-making process, and strategies for integration into routine hospital workflows.

Verified
Journal Article
Open access
PDF file
KBKarima Bahmane
Karima Bahmane
SBSambit Bhattacharya
Sambit Bhattacharya
My Abdelmajid KassemMy Abdelmajid Kassem
My Abdelmajid Kassem
Missense single nucleotide variants (SNVs) represent one of the most common forms of genetic variation and account for a substantial proportion of variants of uncertain significance in clinical databases. Accurate computational classification of these variants remains an important challenge in precision medicine and genomic research. In this study, we present PathoPredictor, an interpretable machine-learning framework designed to distinguish pathogenic from benign missense variants using curated clinical variant data and functional annotations. High-confidence variants were obtained from the November 2023 ClinVar release and annotated using dbNSFP v5.1 (GRCh37). After data filtering, imputation, and normalization, 59,302 expert-reviewed missense variants were retained for model development. Six machine-learning algorithms were evaluated under identical cross-validation conditions applied to the training set. Among the evaluated models, LightGBM demonstrated the strongest overall performance and was selected as the final PathoPredictor classifier, achieving a mean ROC–AUC of 0.93 ± 0.004, accuracy of 0.90 ± 0.006, and Matthew’s correlation coefficient of 0.80 ± 0.008 across five cross-validation folds. Model interpretability was examined using SHAP (SHapley Additive exPlanations), enabling both global feature ranking and variant-level explanation of predictions. Temporal validation using ClinVar variants submitted after November 2023 showed consistent predictive performance on previously unseen submissions within the same database ecosystem (ROC–AUC = 0.91). While the framework demonstrates strong discrimination and structured interpretability, potential limitations include training data bias and partial circularity associated with the inclusion of existing meta-predictors. Overall, PathoPredictor provides a reproducible and interpretable computational framework for integrating functional annotations in missense variant prioritization, supporting research and genomic analysis workflows.

12 June 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
KBKarima Bahmane
Karima Bahmane
ACAlkhalil Brahim Chaouki
Alkhalil Brahim Chaouki
Background: Thyroid nodule segmentation in ultrasound imaging is clinically important but typically requires large volumes of expert annotations. Foundation models offer a promising approach to reducing labeled data requirements, yet the relationship between annotation budget and segmentation performance across adaptation strategies remains insufficiently characterized.Methods: We evaluated four SAM3-based adaptation strategies on TN5000 (5,000 biopsy-confirmed thyroid ultrasound images): full fine-tuning (MedSAM3), zero-shot inference (SAM3-Zero), and Low-Rank Adaptation at ranks 8 and 16 (SAM3-LoRA8, SAM3-LoRA16). All models were initialized via Masked Autoencoder pretraining on the unlabeled corpus and evaluated under four annotation regimes (10%, 25%, 50%, and 100% of labeled data) using Dice, IoU, AUC, sensitivity, specificity, and F1-score, with 95% bootstrap confidence intervals and Bonferroni-corrected significance testing. External validation was conducted on the independent DDTI dataset.Results: At full annotation budget, MedSAM3 achieved the highest internal performance (Dice = 0.815, AUC = 0.989), significantly outperforming the zero-shot baseline (Dice = 0.097, p < 0.001 across Wilcoxon, DeLong, and McNemar tests). At 25% label availability, SAM3-LoRA8 matched or exceeded MedSAM3 (Dice = 0.736 vs 0.726) while updating only 1.85% of model parameters, identifying 25% as the annotation-efficiency crossover threshold. No statistically significant difference was observed between LoRA ranks 8 and 16 on per-image Dice (Wilcoxon p = 0.704), indicating that increasing rank beyond 8 yields no measurable performance benefit for this task. External validation on the DDTI dataset preserved model ranking (MedSAM3 AUC = 0.911); a dissociation between AUC and Dice under domain shift was observed across all fine-tuned models, with implications for cross-site deployment.Conclusions: Full fine-tuning with domain-adapted initialization provides the best segmentation performance at large annotation budgets. LoRA adaptation with rank 8 offers a more annotation-efficient and parameter-efficient alternative, matching full fine-tuning at 25% label availability and requiring approximately half the training time. These findings provide a concrete decision framework for selecting adaptation strategies in annotation-constrained clinical deployments of thyroid ultrasound segmentation.

28 January 2026 in JMIR Preprints
Verified
Preprint
Open access
KBKarima Bahmane
Karima Bahmane
BACKGROUND Artificial intelligence (AI) has reached expert-level performance across many areas of medical imaging, yet this progress has not translated proportionally into improvements in patient outcomes. While deep learning models excel at pixel-level pattern recognition, their impact on clinical decision-making, workflow efficiency, and patient-centered care remains poorly characterized OBJECTIVE This structured narrative review synthesizes evidence from high-quality studies (2018–2025) to evaluate whether imaging AI systems meaningfully improve patient outcomes beyond diagnostic accuracy. The review critically examines clinical integration, workflow implications, ethical considerations, and the persistent gap between algorithmic performance and patient-centered benefit. METHODS A structured search of PubMed, Scopus, IEEE Xplore, and Web of Science (2018–October 2025) identified empirical studies applying AI to human medical imaging and reporting both diagnostic metrics and real-world clinical, workflow, or patient-centered outcomes. Studies were screened independently by two reviewers, and data were extracted using predefined categories : model type, dataset characteristics, validation strategy, performance metrics, workflow impact, patient outcomes, and ethical considerations. RESULTS Ten high-quality studies met the inclusion criteria. Across domains (ophthalmology, mammography, echocardiography, CT, PET/CT, and chest radiography), AI models achieved strong diagnostic performance (pooled mean AUC = 0.91 ± 0.03). However, only 30% of studies reported measurable patient impact and 20% reported workflow improvements. External validation often revealed 5–10% performance degradation, and only four systems were deployed in routine care. Ethical analyses showed emerging concerns regarding bias, explainability, and trustworthiness particularly related to racial inference from imaging data. CONCLUSIONS Medical imaging AI has matured algorithmically but remains clinically immature. Achieving true patient-centered benefit requires shifting from model-centric development to systems-level innovation: multimodal integration, explainable AI, human-in-the-loop designs, equity-aware training, and prospective clinical evaluation. AI will advance from “seeing the organ” to “understanding the patient” only when technical performance aligns with clinical workflows, ethical oversight, and human experience.

12 March 2026 in JMIR Preprints
Verified
Preprint
Open access
KBKarima Bahmane
Karima Bahmane
ICIbrahim Alkhalil Chaouki
Ibrahim Alkhalil Chaouki
BACKGROUND Thyroid nodule segmentation in ultrasound imaging is clinically important but typically requires large volumes of expert annotations. Foundation models offer a promising approach to reducing labeled data requirements, yet the relationship between annotation budget and segmentation performance across adaptation strategies remains insufficiently characterized. OBJECTIVE Thyroid nodule segmentation in ultrasound imaging is clinically important but typically requires large volumes of expert annotations. Foundation models offer a promising approach to reducing labeled data requirements, yet the relationship between annotation budget and segmentation performance across adaptation strategies remains insufficiently characterized. METHODS We evaluated four SAM3-based adaptation strategies on TN5000 (5,000 biopsy-confirmed thyroid ultrasound images): full fine-tuning (MedSAM3), zero-shot inference (SAM3-Zero), and Low-Rank Adaptation at ranks 8 and 16 (SAM3-LoRA8, SAM3-LoRA16). All models were initialized via Masked Autoencoder pretraining on the unlabeled corpus and evaluated under four annotation regimes (10%, 25%, 50%, and 100% of labeled data) using Dice, IoU, AUC, sensitivity, specificity, and F1-score, with 95% bootstrap confidence intervals and Bonferroni-corrected significance testing. External validation was conducted on the independent DDTI dataset. RESULTS At full annotation budget, MedSAM3 achieved the highest internal performance (Dice = 0.815, AUC = 0.989), significantly outperforming the zero-shot baseline (Dice = 0.097, p < 0.001 across Wilcoxon, DeLong, and McNemar tests). At 25% label availability, SAM3-LoRA8 matched or exceeded MedSAM3 (Dice = 0.736 vs 0.726) while updating only 1.85% of model parameters, identifying 25% as the annotation-efficiency crossover threshold. No statistically significant difference was observed between LoRA ranks 8 and 16 on per-image Dice (Wilcoxon p = 0.704), indicating that increasing rank beyond 8 yields no measurable performance benefit for this task. External validation on the DDTI dataset preserved model ranking (MedSAM3 AUC = 0.911); a dissociation between AUC and Dice under domain shift was observed across all fine-tuned models, with implications for cross-site deployment. CONCLUSIONS Full fine-tuning with domain-adapted initialization provides the best segmentation performance at large annotation budgets. LoRA adaptation with rank 8 offers a more annotation-efficient and parameter-efficient alternative, matching full fine-tuning at 25% label availability and requiring approximately half the training time. These findings provide a concrete decision framework for selecting adaptation strategies in annotation-constrained clinical deployments of thyroid ultrasound segmentation.