UFlow-Net: A Unified Approach for Improved Video Frame Interpolation

/in /by

In computer vision, video frame interpolation  plays a significant role in video enhancement by synthesizing intermediate frames to improve temporal resolution and visual quality. This techniques help reduce motion blur, create smoother slow motion videos and enhancing total viewing experience, especially in low frame rate video. This is vital for application like video processing, streaming and video restoration. We are developing UFlow-Net, a deep learning-based model that improves frame interpolation accuracy.

The process starts with a dataset of three consecutive video frames. The first and third frames are used as inputs, while the second frame is used as a reference for evaluation. These frames go through preprocessing, such as resizing, normalizing, and stacking the frames.

Next, the preprocessed frames are passed into UFlow-Net, which consists of two key steps. The Flow-Enhanced Encoder-Decoder captures motion and spatial details from the input frames, and reconstructs the features, keeping the motion consistent. The Refined Frame Synthesis step, refines the features more and generates the missing middle frame by using the learned motion patterns and spatial relationships. We evaluate our model using PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index Measure). Our Model achieved a PSNR of 35.65 dB and SSIM score 0.97. 

Relevant publications: 

F. Israq, S. B. Alam, H. Khatun, S.S. Sarker, S.T. Bhuiyan, M. Haque, R. Rahman and S. Kobashi ” UFlow-Net: A Unified Approach for Improved Video Frame Interpolation” in Proc. 2024 27th International Conference on Computer and Information Technology (ICCIT), Cox’s Bazar, Bangladesh, Dec. 20-22, 2024. 

ROI-Guided Lumbar Spine Degeneration Detection

/in /by

Lumbar Spine Degeneration, otherwise known as Disk Degeneration Disease, is the deterioration/weakening of the intervertebral disks in the lower back, causing them to lose their ability to absorb shock and potentially leading to pain and discomfort due to nerve compression. While lumbar degeneration is a natural part of aging, its progression and severity can vary from person to person. Early detection and classification of the type and stage of degeneration are crucial for effective management and treatment. We are using the power of AI to automate the detection of lumbar degeneration.

Firstly, We merge and pre-process the Magnetic Resonance Imaging (MRI) scans to convert them to RGB images, comprised with multi-label diseases per image.Then, we used an YOLO(You Only Look Once) architecture to detect the region of interests (ROI) that include the intervertebral disks in the MRI scans from both Sagittal and Axial Plane. Finally, the ROIs are classified into different degree of degeneration.

The goal of this research is to detect Lumbar Spine Degeneration as early as possible in order to ensure the best quality of life and outcome for the patient.

Detection-Guided Kidney Abnormality Segmentation

/in /by

Tumors and cysts are two major kidney abnormalities which can lead to cancer if left undetected and untreated timely. The current diagnostic process depend on computed tomography (CT) scan screening which is time-consuming and specialist-dependent. This leads to fatigue for radiologists and doctors. As a result, diagnostic errors increase. To enhance the diagnostic process, we are developing an AI-based automated method for kidney abnormality detection. 

The first step in this method is kidney detection using the YOLOv8 model. This is conducted on 2D sliced images extracted from 3D CT. The detected kidney regions are then cropped to reduce the background area. After that, abnormal region segmentation is conducted using the U-Net segmentation model. The abnormal region consists of either cyst or tumor or both. This produces a mask of abnormal region for each 2D slice. The 2D mask slices are combined to construct a 3D mask of the abnormal region. 

This study aims to automate the detection process of kidney tumor and cyst, offering a faster and enhanced approach to assist the doctors and radiologists in diagnostic process.

Relevant publications:

1. J. Faruk, S. B. Alam, S. S. T. Elma, S. Wasi, R. Rahman and S. Kobashi, “Kidney Abnormality Detection Using Segmentation-Guided Classification on Computed Tomography Images,” 2024 International Conference on Machine Learning and Cybernetics (ICMLC), Miyazaki, Japan, 2024, pp. 414-419, doi: 10.1109/ICMLC63072.2024.10935113.

2. S. Wasi, S. B. Alam, R. Rahman, M. A. Amin and S. Kobashi, “Kidney Tumor Recognition from Abdominal CT Images using Transfer Learning,” 2023 IEEE 53rd International Symposium on Multiple Valued Logic (ISMVL), Matsue, Japan, 2023, pp. 54-58, doi: 10.1109/ISMVL57333.2023.00021

Pelvic Bone Segmentation Guided Fracture Classification

/in /by

Pelvic fractures are critical injuries that require timely and precise diagnosis. We are developing an AI-based automated computer-aided diagnosis (CAD) system to enhance the accuracy and reliability of pelvic fracture detection, ultimately supporting faster and more effective medical assessments.

The first step in our approach is multi-bone segmentation, a deep learning process that identifies and isolates the nine pelvic bones from the X-ray. The segmented bone masks are then aggregated to create a refined mask of the pelvic region. 

Next, we extract the segmented X-ray from the original X-ray using the aggregated mask. This helps the system focus only on relevant areas instead of the entire X-ray. We then feed it into a separate deep learning model for classification. This classification model determines whether a fracture is present.

To improve interpretability, we use GradCAM visualization, which highlights the critical areas the AI focuses during detection. This ensures the model’s decisions are transparent and aligned with anatomical relevance.

Relevant publications:

S. A. Ul Alam, S. Binte Alam, S. Saha, M. Haque, R. Rahman and S. Kobashi, “Pelvic bone region segmentation (PBRS) from X-ray image using convolutional neural network (CNN),” 2023 26th International Conference on Computer and Information Technology (ICCIT), Cox’s Bazar, Bangladesh, 2023, pp. 1-6, doi: 10.1109/ICCIT60459.2023.10441155.