Aryan Neizehbaz

+98 9309129829 · aryanneizehbaz@gmail.com

Education

B.Sc. in Computer Science

Shahid Beheshti University (SBU)

( SBU ranks as the third-best university for computer science in Iran )

CGPA: 3.73 / 4

Sep 2021 - Feb 2025

Selected courses

Neural Network, Dr. Kheradpisheh, 4 / 4
Foundations of Machine Learning, Dr. Farahani, 4 / 4
Artificial Intelligence, Dr. Katanforosh, 4 / 4
Foundations of Data Science, Dr. Kheradpisheh, 4 / 4
Computational Neuroscience, Dr. Kheradpisheh, 4 / 4
Image Processing, Dr. Kheradpisheh, 4 / 4

Honors and Awards

GPA ranked in the top 10% among students in the computer science major.
Awarded full tuition waiver for achieving top 0.3% ranking among 120,000 students in the Iranian university entrance exam.
Accepted at the first level in high school Chemistry Olympiad.

Research Interests

I am passionate about the fields of computer vision and medical imaging, particularly in developing advanced machine learning algorithms to enhance diagnostic accuracy and patient outcomes. I am eager to leverage deep learning techniques to interpret complex medical images, enabling early detection and treatment of diseases. My goal is to integrate machine learning into clinical workflows to provide personalized and precise healthcare solutions. Additionally, I am enthusiastic about exploring the use of artificial intelligence in health informatics to improve the overall efficiency of healthcare systems and contribute to the advancement of medical technology.

Journals

Transforming [177Lu]Lu-PSMA-617 Treatment Planning: Machine Learning-Based Radiodosiomics and Swin UNETR using Pretherapy PSMA PET/CT [PAPER] [CODE]

Elmira Yazdani, Aryan Neizehbaz, Najme Karamzade-Ziarati, Farshad Emami, Habibeh Vosoughi, Mahboobeh Asadi, Atefeh Mahmoudi, Mahdi Sadeghi, Saeed Reza Kheradpisheh, Parham Geramifar

(Medical Physics 52 (10): e70030, October 2025)

Background: Personalized pretreatment dosimetry planning is crucial for optimizing [¹⁷⁷Lu]Lu-prostate-specific membrane antigen-617 (Lu-PSMA) radioligand therapy (RLT) in metastatic castration-resistant prostate cancer (mCRPC) patients.

Purpose: This study addresses two goals. First, we develop a machine learning (ML)-based pretreatment planning model to predict post-therapy absorbed doses (ADs) in metastatic lesions by integrating clinical biomarkers (CBs) with radiomic features (RFs) and dosiomic features (DFs) extracted from [⁶⁸Ga]Ga-PSMA-11 (Ga-PSMA) PET/CT to enhance predictive accuracy. Second, we develop a transformer-based deep learning (DL) architecture to predict Monte Carlo (MC)-derived dose rate maps (DRMs), minimizing reliance on computationally intensive MC simulations.

Methods: For the ML objective, retrospective posttreatment dosimetry data from 20 mCRPC patients treated with Lu-PSMA RLT were used as ground truth labels. Patient-specific MC dosimetry was employed on Ga-PSMA PET/CT images using the GATE v9.1 toolkit to generate DRMs. After image preprocessing, RFs and DFs were extracted from Ga-PSMA CT images and DRMs using LIFEx v7.4.0. Feature selection was conducted via recursive feature elimination with Benjamini-Hochberg correction (q<0.05). Multiple nonlinear regression models were trained using leave-one-out crossvalidation (LOOCV), and model interpretability was assessed using SHAP and LIME radar plots. A shifted windows UNET Transformers (Swin UNETR) architecture with selfsupervised learning (SSL) pretraining was employed to predict voxel-wise PET-based DRMs for the DL objective. The model was fine-tuned on MC-labeled DRM data from 30 patients (including 10 additional cases) using 5-fold cross-validation.

Results: The ensemble tree regressor (ETR) using selected CT RFs, PET DFs, and significant CBs achieved R² = 0.82 and RMSE = 0.67 Gy/GBq. For DRM prediction, the SSL-pretrained Swin UNETR achieved R² of 0.97, NRMSE of 0.003 Gy/GBq, and a Gamma pass rate of 99.08%, closely matching MC-derived DRMs.

Conclusions: Integrating ML-based radiodosiomics and transformerbased DL enables accurate, efficient lesion AD and DRM prediction from pretherapy PET/CT, supporting personalized Lu-PSMA RLT planning.

Explainable artificial intelligence for pneumonia classification: Clinical insights into deformable prototypical part network in pediatric chest x-ray images [PAPER]

Elmira Yazdani, Aryan Neizehbaz, Najme Karamzade-Ziarati, Saeed Reza Kheradpisheh

(Journal of Medical Imaging and Radiation Sciences 56 (5): 102023, September 2025)

Background: Pneumonia detection in chest X-rays (CXR) increasingly relies on AI-driven diagnostic systems. However, their “blackbox” nature often lacks transparency, underscoring the need for interpretability to improve patient outcomes. This study presents the first application of the Deformable Prototypical Part Network (DProtoPNet), an ante-hoc interpretable deep learning (DL) model, for pneumonia classification in pediatric patients’ CXR images. Clinical insights were integrated through expert radiologist evaluation of the model’s learned prototypes and activated image patches, ensuring that explanations aligned with medically meaningful features.

Methods: The model was developed and tested on a retrospective dataset of 5,856 CXR images of pediatric patients, ages 1–5 years. The images were originally acquired at a tertiary academic medical center as part of routine clinical care and were publicly hosted on a Kaggle platform. This dataset comprised anterior-posterior images labeled normal, viral, and bacterial. It was divided into 80 % training and 20 % validation splits, and utilised in a supervised five-fold cross-validation. Performance metrics were compared with the original ProtoPNet, utilising ResNet50 as the base model. An experienced radiologist assessed the clinical relevance of the learned prototypes, patch activations, and model explanations.

Results: The D-ProtoPNet achieved an accuracy of 86 %, precision of 86 %, recall of 85 %, and AUC of 93 %, marking a 3 % improvement over the original ProtoPNet. While further optimisation is required before clinical use, the radiologist praised D-ProtoPNet’s intuitive explanations, highlighting its interpretability and potential to aid clinical decision-making.

Conclusion: Prototypical part learning offers a balance between classification performance and explanation quality, but requires improvements to match the accuracy of black-box models. This study underscores the importance of integrating domain expertise during model evaluation to ensure the interpretability of XAI models is grounded in clinically valid insights.

From Bitewing to Report: VLM-Assisted Pulpal Exposure Classification and Risk Stratification

(in preparation)

Developing a vision-language model that leverages planar thyroid images and clinical data to generate diagnostic descriptions, using doctors' reports as labeled outputs to enhance accuracy in thyroid disease analysis.

Conferences

Employing Enhanced ResNet-Based Architecture for Automatic Segmentation of Tumoral Lesions and Organs on [68 Ga]Ga-PSMA11 PET/CT Images in Prostate Cancer [ABSTRACT]

Elmira Yazdani, Aryan Neizehbaz, Najme Karamzade-Ziarati, Habibeh Vosoughi, Mahdi Sadeghi, Saeed Reza Kheradpisheh, and Parham Geramifar

(Presented at the 13th Iranian Congress of Medical Physics)

Introduction: Prostate-specific membrane antigen (PSMA) PET/CT has emerged as a powerful tool for quantitative image analysis, particularly in the context of radioligand therapy (RLT) for metastatic castration-resistant prostate cancer (mCRPC). However, the manual segmentation of these images remains labor-intensive and prone to errors, limiting its scalability in clinical practice. Recent advancements in deep learning offer a promising solution, enabling faster and more accurate segmentation that reduces dependency on specialized expertise, thereby improving clinical workflow and efficiency.

Methods: In this study, we propose an improved ResNet architecture designed to segment lesions and organs at risk (OARs), including salivary glands, liver, spleen, kidneys, bowel, and bladder, in 100 [68 Ga]Ga-PSMA PET/CT images. To handle image variability, normalization and data augmentation techniques were applied. The model, trained over 1000 epochs, used a residual convolutional block to extract deep features from the preprocessed images. For better segmentation accuracy, features from all ResNet layers were combined into a unified output that merged deep semantic and shallow appearance features. The model’s performance was evaluated using a 5-fold cross-validation, with 80% of the data used for training and 20% for testing. Ground truth segmentation was provided by an expert clinician using 3D Slicer, and the models were implemented using PyTorch and MONAI on an NVIDIA Tesla k80 GPU.

Results: The improved ResNet achieved an average Dice similarity coefficient of 0.64 for lesions and 0.82 for OARs, with corresponding precision values of 0.60 and 0.77. Recall values were 0.76 for lesions and 0.91 for OARs, outperforming the benchmark U-Net model. This enhanced ResNet architecture, with its encoder-decoder structure, successfully captured multi-scale contextual features, enabling precise and rapid target identification. The automated segmentation technique presented in this study shows great potential for advancing clinical applications, such as personalized dose prediction in RLT.

Experiences

Teaching

Artificial Intelligence, instructed by Dr. Ali Katanforosh

Designed mini projects and coding assignments.
Mentored students

An Introduction to Artificial Intelligence, instructed by Dr. Atefeh Aghaee

Designed coding assignments and Quizzes.
Mentored students

Advanced Programming, instructed by Dr. Saeedreza Kheradpisheh

Mentored students and provided assistance with assignment and coding problems.
Created and designed problems for student assignments to reinforce OOP principles.

Front End Developer

Shahid Beheshti University Science and Technology Park

Projects

Gold Trading System

Online gold buying and selling system and the possibility of storing gold. Buy and sell gold at the spot price. It has an admin and management panel for registering prices and conﬁrming purchases and sales, the possibility of checking and reporting on purchases and sales and online payment. It has the possibility of users chatting with admins and asking questions with them.

Dormitory Managment System

Developed the front-end of a dormitory management system using React. This project involved creating and managing a range of components to facilitate efficient dormitory operations. Responsibilities included implementing user authentication, secure access, and handling various administrative functions such as personnel management, inventory tracking, payment history, and request handling. The goal was to enhance the overall efficiency and usability of the dormitory management system through a well-structured and intuitive user interface.

Sep 2022 - Mar 2023

IT Specialist Intern

Global Village Computer Store and Services

Projects

Diagnosing and resolving hardware and software issues
Conducting technical repairs and maintenance
Providing client support and system troubleshooting
Assisting in the deployment and implementation of new technologies
Ensuring optimal system performance and reliability

Jan 2025 - Aug 2025

Projects

Skin Cancer Classification

Developed a comprehensive machine learning model for skin cancer classification using the HAM10000 dataset, incorporating data augmentation, Random Forest, SVM, and CNN models for accurate diagnosis and analysis.

Large Language Model Text Generator for Persian Wikipedia

Developed an LLM for text generation using the Persian Wikipedia dataset, covering dataset creation, model design, training, evaluation (Perplexity, ROUGE), and text generation challenges.

Reinforcement Learning for BipedalWalker-v3

Implemented a reinforcement learning algorithm to train an agent for bipedal robot control, including environment setup, training, and performance evaluation.

Click-Through Rate Prediction Using Machine Learning

Developed a model to predict ad click likelihood, involving EDA, data cleaning, feature engineering, and algorithm implementation, with performance evaluated through cross-validation.

Autoencoder Reconstruction of Mixed MNIST and CIFAR-10 Images

Built an autoencoder in PyTorch to reconstruct images from combined MNIST and CIFAR-10 datasets, involving image preprocessing and quality evaluation.