Amirmohammad MohammadiHomepage

Research

01 / Computer vision & digital health

Vision-language models for digital pathology.

I evaluate a pretrained computational pathology vision-language model for predicting ethanol drinking status from nonhuman primate liver whole-slide images with few labeled examples, comparing visual adapters and textual task residuals with attention-based multiple instance learning. In four-shot experiments, visual adaptation achieved 78.1% mean balanced accuracy versus 43.5% for the ABMIL baseline, using subject-disjoint evaluation over 30 runs.

Method figure showing tissue patches, frozen image and text encoders, multiple instance learning, a visual adapter, task residuals, and cosine-similarity prediction.
Whole-slide classification with frozen vision-language encoders and trainable adaptation modules. View full-size figure.

02 / Foundation-model adaptation

Distribution-aware adaptation for classification.

I developed and integrated a learnable histogram module into pretrained Transformers to capture target-domain feature distributions while keeping the backbone frozen. The method enables parameter-efficient adaptation for passive and active sonar classification, with comparisons of accuracy, trainable parameters, and feature representations. On VTUAD, it achieved 91.8% accuracy versus 89.8% for the conventional adapter baseline.

A trainable histogram module runs alongside frozen Transformer components.
Histogram-based tuning within a Transformer block. View full-size figure.

03 / Physiological time series

Physics-informed models for physiological prediction.

I co-developed physics-informed neural networks for cuffless blood-pressure estimation from physiological time series. The approach incorporates physiological relationships expressed through Taylor approximations into supervised training. The collaborative study reported an average 15 times reduction in ground truth training labels compared with the tested time-series baselines.

A neural network combines physiological features with supervised and physics-based training objectives.
Combining physiological relationships and measured data during training. View full-size figure.

04 / Biomedical sensing

An integrated wearable for stress detection.

During my master’s research, I designed an ECG/EDA wearable stress-detection system, from its circuitry and Bluetooth Low Energy electronics to supervised learning algorithms. The 18-participant study achieved 94% classification accuracy and threefold battery life relative to the comparison devices.

Wearable electronics including ECG and electrodermal sensing, power management, and a Bluetooth Low Energy microcontroller.
An integrated physiological sensing platform. View full-size figure.

Current projects

Computer vision evaluation & multimodal learning.

Computer vision evaluation

I contribute to a modular evaluation pipeline for image classification, semantic segmentation, and object detection, using shared protocols to compare pretrained models and adaptation strategies.

Multimodal acoustic classification

My dual-encoder work combines waveform and spectrogram representations through differentiable Choquet-integral fusion, learning how to combine complementary information for classification.

Read the preprint