Publications

2025

1. An efficient Swin Transformer Architecture for Cutaneous Melanoma Detection

   Jan 2025 - Present

   Reza Esmkhani, Milad Soleimani, Soroush Ziaee, Maliheh Sabeti, Nava Eslami
   
   Melanoma is the deadliest and most unpredictable type of skin cancer. This paper aims to develop a robust image classification model for the early detection of melanoma, a deadly skin cancer. This investigation addresses the limitations of previous methods by developing and validating parameter-efficient Swin Transformer adapters, which are explicitly optimized for melanoma classification. Our approach implements bottleneck adapters with down-projection, non-linearity, up-projection architecture, requiring only 2-4% of total model parameters compared to LoRA's 0.1-0.3%, while providing superior modularity for multi-task scenarios despite introducing up to 50% throughput degradation in sequential processing. The resulting framework is designed to enable deployment on resource-constrained platforms while maintaining clinical-grade accuracy, thereby facilitating broader access to AI-assisted diagnostic. The methodology encompasses stratified 4-fold cross-validation on established benchmark datasets, measuring sensitivity, specificity, F1-score, and computational metrics including inference time (target <100ms), memory usage (within 4GB constraints), and FLOPs reduction compared to full fine-tuning. This work advances current understanding of evidence supporting parameter-efficient approaches for medical AI while providing a validated framework for clinical deployment.
   [Manuscript]

2021

1. A Deep Learning Algorithm for Classifying Grasp Motions using Multi-session EEG Recordings

   Mar 2020 - Jan 2021

   Andy Partovi, Seyed Mehrshad Hosseini, Milad Soleymani, Kiana Liaghat, Soroush Ziaee, Erfan Habibi Panah Fard
   2021 9th International Winter Conference on Brain-Computer Interface (BCI)
   
   The classification of motor imagery tasks using scalp EEG signals is a complicated procedure in BCI especially when the task comprises multiple gestures of the same hand. In this paper, we present a classification method to distinguish three grasp motion classes (cylindrical, spherical, and lumbrical) of one hand over two-day training sessions in 15 subjects in a public dataset. We have developed two ensemble methods consisting of (anomaly detection + fully connected neural network) and (anomaly detection + convolutional neural network) to classify grasp motion and have achieved more than 80% classification accuracy in 3 subjects and an average accuracy of 57% among the full cohort. Our results confirm the possibility of utilizing neural networks to decode motor movement intentions from scalp EEG in a complicated task.
   [IEEE Xplore]