Full resolution convolutional neural network based organ and surgical instrument classification on laparoscopic image data

• Published in: Biomedical signal processing and control 2024-01, Vol.87, p.105533, Article 105533
• Main Authors: SR Konduri, Praveen, Siva Nageswara Rao, G.
• Format: Article
• Language: English
• Subjects: Average spectrum; Deep learning; Full convolution neural network; Laparoscopic image; Local binary pattern
• ISSN: 1746-8094
• DOI: 10.1016/j.bspc.2023.105533

•To present a deep learning based architecture known as Full resolution Convolutional Neural Network (FrCNN) for efficient organ and instrument classification using a laparoscopic image database.•The proposed FrCNN based organ classification aims to efficiently examine the organs inside the abdomen without making large incisions in human skin.•To automatically identify the presence of surgical instruments in laparoscopic videos that intends to detect the tool used at each time in surgery. This work focuses on classifying different organs like the uterus, ovary, oviduct, liver, and colon as well as various surgical instruments such as bipolar, scissors and grasper with the help of two publically available databases such as ITEC LapGyn4 and Cholec80. The different procedures carried out in this methodology are pre-processing for noise removal, clustering using an improved K means algorithm, texture and spectral feature extraction, suitable feature selection using an improved cuckoo search optimization algorithm, and finally, classification. Here utilize an efficient deep learning architecture for achieving efficient classification, known as Full Resolution Convolutional Neural Network (FrCNN). This high-performance neural network architecture is formed by replacing conventional CNN’s downsampling layer with a full resolution layer. Furthermore, the proposed technique’s results are evaluated using different parameters such as accuracy, precision, f1-score, recall, etc. The accuracy, precision, recall, and f1-score generated for Dataset 1(LapGyn4) are 99.743 %, 98.650 %, 99.192 %, and 98.914 %, respectively. Similarly, instrument classification gained outcomes as 99.275%, 98.686 %, 98.973 %, and 98.827 %, respectively. The organ classification outcomes obtained in dataset 2 (Cholec80) are 99.452 %, 98.907 %, 98.904 %, and 98.905 %, respectively. For the Cholec80 dataset, the instrument classified accuracy is 98.826 %, precision is 98.822 %, recall is around 98.822 %, and the f1-score is 98.822 %. For dataset 3, the accuracy achieved is 99.34% and 99% for organ and instrument classification. The overall results indicate that the proposed approach is better than all other existing methods.