Deep learning-based surgical phase recognition and interpretable workflow anomaly detection in laparoscopic cholecystectomy.
| dc.contributor.author | Opio, Jonathan | |
| dc.date.accessioned | 2026-09-24T07:02:51Z | |
| dc.date.available | 2026-09-24T07:02:51Z | |
| dc.date.issued | 2026 | |
| dc.description | Dissertation | |
| dc.description.abstract | This thesis presents an interpretable deep learning approach for analysing surgical workflows and identifying statistically unusual patterns in laparoscopic cholecystectomy videos. The proposed approach uses a two-stage architecture that combines deep learning-based surgical phase recognition with probabilistic normative modelling. In the first stage, a multimodal temporal architecture is developed by integrating ResNet-50 visual features, instrument information derived from the CholecT50 annotations, a Bidirectional Gated Recurrent Unit (BiGRU) encoder, and a temporal attention mechanism. The model recognises surgical phases from short video windows together with instrument signals. Using a 16-frame sliding window and class-weighted cross-entropy loss, the selected attention-based model achieved a validation accuracy of 74.16%. The selected model was subsequently evaluated on the held-out test set and achieved an accuracy of 72.43%, which provides the principal measure of model generalisation. In the second stage, two normative models are learned from the training data to represent statistically common workflow patterns. The first captures phase-transition probabilities to characterise the expected temporal ordering of surgical phases, while the second captures phase-instrument relationships to characterise commonly observed instrument usage patterns. Deviation thresholds are calibrated from the corresponding training distributions using percentile-based procedures. The transition and semantic components use lower-tail thresholds, while the combined window-level workflow-deviation score uses a 95th-percentile threshold. On the Stage 2 validation data, the approach flagged 4.86% of validation records when ground-truth phase labels were used and 7.44% of validation records when predicted phases were used. These results describe the behaviour of the statistical flagging procedure under the calibrated thresholds and provide empirical evidence that phase-recognition errors can propagate into subsequent workflow-deviation flagging. For each flagged workflow event, the approach provides structured information including the video identifier, frame number, deviation score, component-wise flags, and an interpretable explanation such as a rare phase transition or an expected instrument being absent. These explanations decompose the statistical deviation into temporal and semantic components, supporting transparent inspection of the model’s findings and subsequent expert review. Importantly, the proposed approach identifies statistically unusual workflow patterns relative to patterns learned from the training data. Flagged events are candidates for expert review and should not be interpreted as clinically verified surgical errors, complications, negligence, misconduct, or other forms of wrongdoing. Clinical and domainexpert validation is required before a flagged event can be considered a genuine clinical or procedural anomaly. The approach is therefore intended to support post-hoc workflow analysis, structured auditability, and forensic review rather than to independently establish clinical or legal conclusions | |
| dc.description.sponsorship | Dr. Godliver, Owomugisha ; Prof. Gilbert , Gilibrays Ocen ; Busitema University | |
| dc.identifier.citation | Opio, J. (2026). Deep learning-based surgical phase recognition and interpretable workflow anomaly detection in laparoscopic cholecystectomy. [Unpublished dissertation]. Busitema University. | |
| dc.identifier.uri | https://bdears.busitema.ac.ug/handle/123456789/9632 | |
| dc.language.iso | en | |
| dc.publisher | Busitema University | |
| dc.title | Deep learning-based surgical phase recognition and interpretable workflow anomaly detection in laparoscopic cholecystectomy. | |
| dc.type | Other |