Deep Neural Networks in the Diagnosis of Postoperative Complications of Acute Appendicitis

Introduction

Digital models are needed in medicine for diagnosis and prediction. Such models are especially needed in personalized medicine. In this area, it is necessary to evaluate and predict the patient's condition from a priori knowledge obtained from other patients. Therefore, a new direction appeared – predictive medicine (Brigham and Johns, 2012;Miner, et al., 2014). Predictive medicine, or “in silico medicine” is the use of computer modeling and intelligent technologies in the diagnosis, treatment and prevention of diseases. Using predictive medicine, the doctor can determine the likelihood of the development of certain diseases and choose the optimal treatment. Predictive medicine begins to be applied in surgery (Joskowicz, 2017). The prognosis in surgery consists in the preoperative evaluation of various surgical interventions and in the evaluation of possible outcomes of surgical interventions.

This chapter discusses the diagnosis of postoperative complications of acute appendicitis. The problems of early diagnosis, treatment, prevention and prognosis of complications of acute appendicitis are relevant. Among urgent operations, the proportion of appendicitis (appendectomy) removal operations is about 85%. Despite the active use of medical equipment for ultrasound, computer, magnetic resonance, laparoscopic, endoscopic diagnosis of acute appendicitis and its complications, the problem remains unsolved. The solution to this problem is the ability of the doctor to apply diagnostic methods and to objectively interpret them. It is necessary to perform an appendectomy with a clinical analysis of the situation in the abdominal cavity in a timely manner and on the basis of indications: to identify infiltrates, abscesses, local peritonitis. It is necessary to develop treatment tactics that prevent the development of purulent-inflammatory complications. It should be noted that purulent-inflammatory complications after appendectomy occur in 2.7% –39.1% of patients. So far, mortality in acute appendicitis varies from 0.1% to 1.6%. Solving these problems requires the development of methods for predicting postoperative appendicitis complications. In (Prabhudesai, et al., 2008;Park and Kim, 2015), artificial neural networks that diagnose the presence of acute appendicitis were developed. In (Juliano, et al., 2017), a study of risk factors associated with complications of acute appendicitis is conducted, specific types of complications were investigated in (Bakti, et al., 2011). However, mathematical models to predict postoperative appendicitis complications are absent.

Experiments conducted by the authors with a three-layer neural networks for the direct distribution of the classical architecture (Haykin, 2008) showed that such networks do not provide high diagnostic accuracy. In addition, such networks are prone to overfitting. Currently, neural networks of deep architecture, using specific learning algorithms and activation functions, are very popular (Goodfellow, et al., 2016;Aggarwal, 2018). Such networks are used to solve various problems of image processing, sound sequences, etc. Their main advantage is that they are able to capture very complex non-linear relationships in the data. The tasks of medical diagnostics, as a rule, have a large number of non-linearly interconnected signs characterizing the patient's condition. Moreover, the sets of features that characterize patients with different conditions may differ quite slightly. Therefore, it seems promising to use neural networks of deep architecture for medical diagnostics, in particular, for the diagnosis of postoperative appendicitis complications.

The aim of the work is the study of direct propagation neural networks of deep architecture for the diagnosis of postoperative appendicitis complications.