Hi there, I’m Amirhossein Eskorouchi!

Description: Conventional object detectors optimize individual detections but may fail to recover the physical size distribution needed for manufacturing quality decisions. UOT-DETR addresses this problem by adding an unbalanced optimal transport regularizer to DETR training, encouraging the predicted wood-chip population to preserve the reference size distribution while retaining the standard DETR inference process.

Outcome: UOT-DETR achieved an mAR50 of 95, an overall mAR50-95 of 72, and a small-object mAR50-95 of 42 on dense real-world wood-chip imagery. Across ten paired experimental seeds, it improved small-object mAR50 by 24.7 percentage points and small-object mAP50 by 15.4 percentage points over DETR, with the largest gains occurring in medium- and high-clutter scenes. The training-time regularizer introduces no inference-time overhead.

Project support: USDA-NIFA

Description: Extracapsular extension, the spread of cancer beyond a lymph-node capsule, is a major prognostic factor in head and neck cancer, but its preoperative identification on CT is difficult because the imaging signs are subtle and pathology-confirmed labels are scarce. Rather than learning from the entire scan or using segmentation only as preprocessing, this framework embeds lymph-node anatomy directly into both feature learning and decision-making. A dual-output 3D SwinUNETR jointly localizes nodal and perinodal regions and classifies ECE, while the predicted mask forms soft spatial weights that determine which latent features reach the classifier. Clinically informed mask expansion models disease spread beyond the nodal boundary, and Mask-in-Mask pretraining on approximately 1,500 unlabeled CT volumes improves learning under limited supervision.

Outcome: On 150 patients with pathology-confirmed ECE status, the proposed model achieved an accuracy of 0.924, an AUC of 0.928, a sensitivity of 0.926, and a segmentation Dice score of 0.821. It achieved the strongest accuracy, AUC, sensitivity, and Dice among the evaluated methods. At inference, the model predicts its own anatomical region and uses that prediction to guide classification, preserving a fully automatic pathway while grounding decisions in clinically meaningful evidence.

Project support: NIH-USA

Description: Conventional wood-chip quality assessment relies heavily on periodic sampling and offline measurements. Wood-Chip Monitor moves this process closer to real-time operations through a portable NVIDIA Jetson and TensorRT platform that combines dense chip detection, physical size estimation, size-distribution monitoring, oversize alerts, and image-based moisture assessment. A local dashboard converts the model outputs into rolling statistics, configurable quality rules, event records, and live visual feedback for operators.

Outcome: The project translates research models into a complete edge-AI prototype spanning image acquisition, accelerated inference, physical measurement, quality analytics, backend services, and an operator-facing interface. The geometry engine achieved an mAR50 of 95% with a reported size-distribution error of KL = 0.04, while the lightweight moisture engine achieved 91% classification accuracy.

Project support: USDA-NIFA

Description: Plant health management requires continuous context rather than isolated sensor readings. PlantLife365 connects an ESP32 multisensor node with an authenticated Django platform to collect temperature, humidity, soil-moisture, light, gas, and optional image data. The platform organizes these measurements into live conditions, historical trends, configurable alerts, rolling statistics, and exportable user-owned records.

Outcome: The project translates multimodal sensing into an operational decision-support workflow spanning edge acquisition, secure ingestion, user-scoped data services, browser-based monitoring, exploratory regression analytics, and an optional local AI assistant.

Project support: Mississippi Institutions of Higher Learning

Description: This study investigates sleep-stage-specific physiological connectivity as an objective source of ADHD biomarkers. Pearson correlations across 17 EEG, EOG, EMG, and ECG channels are converted into weighted graphs for each participant and sleep stage. Average shortest-path length summarizes each graph, forming a five-dimensional participant representation spanning Wake, Sleep Stages 1, 2, and 3-4, and rapid eye movement sleep. A Random Forest classifier then evaluates these graph-derived features through nested cross-validation.

Outcome: The published study analyzed 48 participants and more than 40,000 30-second PSG epochs. The model achieved average accuracy of 72%, precision of 71%, recall of 85%, and an F1 score of 76%. Feature importance and permutation analysis identified Sleep Stage 1 and Sleep Stage 3-4 as the most influential features for distinguishing ADHD and control participants.

Project support: System Intelligence and Optimization Lab