Bioinformatics is an interdisciplinary field that combines biology, computer science, and information technology to analyze and interpret biological data. As the emergence of deep learning has transformed this field, providing unparalleled insights into complex biological data, we explore state-of-the-art models to unravel the complex patterns in biological signals, including but not limited to Electroencephalograms (EEGs), Electrocardiograms (ECGs), and functional Magnetic Resonance Imaging (fMRI). Our work is focused on developing both predictive and generative models within this domain.
Potential topics:
I. Medical Image Analysis & Diagnosis (Strong Bioinformatics Focus)
- Segmentation/Classification of Thyroid Slides with Explainable AI:
Description: Develop and compare deep learning models (e.g., CNNs, Transformers) for automated segmentation of thyroid tissue in histopathology slides, classifying regions as benign or malignant. Crucially, focus on explainability. Use techniques like Grad-CAM, SHAP values, or attention mechanisms to understand why the model makes its decisions. This is vital for clinical trust and validation. - Review on Medical Diagnosis using Electromagnetic Waves – Focusing on Machine Learning Integration:
Description: This is a review paper, but it shouldn’t just be descriptive. Focus the review on how machine learning techniques are being used to improve the interpretation of data from electromagnetic wave-based medical diagnostics (e.g., EEG, MEG, fMRI, microwave imaging). Identify gaps in current research and potential future directions. - Segmentation/Classification on Various Types of Images (X-ray, MRI, Ultrasound, etc.) for Various Organs – with a Focus on Domain Adaptation:
Description: Instead of just building separate models for each organ and modality, investigate domain adaptation techniques. Train a model on one dataset (e.g., chest X-rays) and adapt it to perform well on another (e.g., brain MRIs). This addresses the problem of limited labeled data in some medical domains.
II. Computer Vision & Feature Engineering (Strong CS Focus)
- Low-Level Visual Features – Robustness to Noise and Adversarial Attacks:
Description: Investigate the robustness of different low-level visual features (e.g., SIFT, HOG, color histograms) to various types of noise and adversarial perturbations. Develop methods for making these features more resilient. - Gestalt Features Implementation & Evaluation for Scene Understanding:
Description: Implement algorithms to extract Gestalt features (e.g., proximity, similarity, closure, continuity) from images. Evaluate their effectiveness in tasks like object recognition and scene understanding. This is about modeling how humans perceive visual scenes. - Low-level visual features generalization against humans on objectnet :
Description: Evaluate how well low-level visual features correlate with human perception on the ObjectNet dataset (which contains human annotations for various attributes and relationships). This is about bridging the gap between computer vision and human vision.
III. Advanced Network Architectures & Optimization (Mix of Bioinformatics/CS)
- Dual Path Network for Compression/Super Resolution – with Perceptual Loss Functions:
Description: Explore dual-path networks (DPNs) for image compression or super-resolution. Key: Use perceptual loss functions (based on features extracted from pre-trained deep learning models) to improve the visual quality of the compressed/super-resolved images, making them more visually appealing and less prone to artifacts. - Multi-Teacher for Panoptic Segmentation – Focusing on Uncertainty Estimation:
Description: Implement a multi-teacher learning framework for panoptic segmentation (segmenting both objects and their boundaries). Crucially, focus on uncertainty estimation. Use the ensemble of teachers to quantify the confidence of the predictions, which is important in safety-critical applications like medical imaging. - Neuroscience-based Network Optimisation via RL – Exploring Sparse Connectivity:
Description: Use principles of neuroscience (e.g., synaptic pruning, sparse connectivity) to guide the optimization of deep learning networks using reinforcement learning (RL). The goal is to create more efficient and robust models.
Involved Researchers:
- Alessio Fagioli