Presenter(s)
Andy Yan
Abstract
Carcinogenic changes are known to influence both the timing and progression of mitosis in colorectal cancer cells, which makes quantitative analysis of mitotic events an important but challenging problem. In DNA-only time-lapse microscopy, these events tend to be short-lived, visually inconsistent, and relatively sparse, so methods based on dense segmentation are often not a good fit, especially when pixel-level annotations are limited. In this project, we developed a spatio-temporal CNN-based method to detect mitotic events in DNA-only time-lapse microscopy. Mitosis detection was treated as an event localization problem rather than a segmentation task. Each event was represented by its center, and Gaussian heatmaps were used as supervision instead of full masks. A lightweight U-Net-based architecture was trained on 256×256 cropped patches. To include temporal information, consecutive frames were stacked together as input, which provided some context without making the model overly complex. During inference, local maxima detection and non-maximum suppression were applied to the predicted heatmaps to obtain event coordinates, and detections were linked across frames to form trajectories. This also made it possible to estimate mitotic duration. Model performance was sensitive to the choice of Gaussian scale, with σ = 3 and r = 8 giving the best overall results, where σ represents the spread of the Gaussian target and r represents the radius of the supervised region. Using temporal input also helped, and T = 3 gave the most stable results overall, where T denotes the number of consecutive frames used as input. Overall, this approach reduced annotation effort and demonstrates a scalable approach for quantitative analysis of mitotic dynamics.
College
College of Science & Engineering
Department
Computer Science
Campus
Winona
First Advisor/Mentor
Mingrui Zhang; Trung Nguyen
Location
Kryzsko Great River Ballroom, Winona, Minnesota; United States
Start Date
4-23-2026 9:00 AM
End Date
4-23-2026 10:00 AM
Presentation Type
Poster Session
Format of Presentation or Performance
In-Person
Session
1a=9am-10am
Poster Number
75
Spatio-Temporal Detection of Mitotic Events Using a CNN-Based Approach
Kryzsko Great River Ballroom, Winona, Minnesota; United States
Carcinogenic changes are known to influence both the timing and progression of mitosis in colorectal cancer cells, which makes quantitative analysis of mitotic events an important but challenging problem. In DNA-only time-lapse microscopy, these events tend to be short-lived, visually inconsistent, and relatively sparse, so methods based on dense segmentation are often not a good fit, especially when pixel-level annotations are limited. In this project, we developed a spatio-temporal CNN-based method to detect mitotic events in DNA-only time-lapse microscopy. Mitosis detection was treated as an event localization problem rather than a segmentation task. Each event was represented by its center, and Gaussian heatmaps were used as supervision instead of full masks. A lightweight U-Net-based architecture was trained on 256×256 cropped patches. To include temporal information, consecutive frames were stacked together as input, which provided some context without making the model overly complex. During inference, local maxima detection and non-maximum suppression were applied to the predicted heatmaps to obtain event coordinates, and detections were linked across frames to form trajectories. This also made it possible to estimate mitotic duration. Model performance was sensitive to the choice of Gaussian scale, with σ = 3 and r = 8 giving the best overall results, where σ represents the spread of the Gaussian target and r represents the radius of the supervised region. Using temporal input also helped, and T = 3 gave the most stable results overall, where T denotes the number of consecutive frames used as input. Overall, this approach reduced annotation effort and demonstrates a scalable approach for quantitative analysis of mitotic dynamics.
Comments
Yan, Andy