Fuzzy Inference System for Pioped-compliant Diagnosis of Pulmonary Embolism

A Fuzzy Inference System was developed to aid in the diagnosis of Pulmonary Embolism using ventilation-perfusion scans and correlated chest x-rays. The diagnosis achieved needed to be accurate and reliable, comparable to that of a nuclear medicine radiologist. The Mamdani fuzzy model has been employed to implement the inference system. Sources of expertise included the criteria defined by PIOPED, Biello, and two nuclear radiologists at the Medical College of Ohio, USA. The proposed Fuzzy Inference System requires the physician to extract feature information from the ventilation-perfusion scans and correlating chest x-rays to facilitate the computer-aided diagnosis process. The proposed inference system has been subjected to testing at various levels including actual user testing. Testing results are encouraging to warrant further exploration of the proposed technique as an integral component in a fully automated diagnosis system. INTRODUCTION Artificial Intelligence has, increasingly and successfully, been applied to a variety of problems in medicine during the last few decades. Artificial Intelligence embodies various techniques including Fuzzy Inference Systems, Artificial Neural Networks, Knowledge-Based Systems, and Bayesian Belief Networks. The main research goal is to address and diagnose for a disease in medicine known as Pulmonary Embolism. The salient features of Fuzzy Logic make it ideal to be used in the field of medical diagnosis. A radiologist’s thought process in arguing a diagnosis for Pulmonary Embolism, if analyzed, indicates that fuzzy reasoning takes place, which can be easily modeled by a Fuzzy Inference System. The initial attempts to diagnose for Pulmonary Embolism employed Knowledge Based Systems [Burton et al. 1984; Ezquerra et al. 1986; Datz et al. 1994; Gabor et al. 1994]. Later attempts employed Artificial Neural Networks [Banish et al. 1993; Tourassi et al. 1995; Fisher et al. 1996] and Bayesian Belief Networks [Bodas et al. 1999] to improve on the diagnosis. Above techniques effectively demonstrated the feasibility of addressing this important medical problem using Artificial Intelligence algorithms. The proposed research attempts to address some of the difficulties faced by the previous researchers/methods. FUZZY INFERENCE SYSTEM FOR PULMONARY EMBOLISM The proposed Fuzzy Inference System (FIS) for diagnosis of Pulmonary Embolism employs the criteria professed by the PIOPED investigators and those of Biello’s. The diagnostic software tool developed through this research follows the (modified) PIOPED criteria mainly as it is considered the standard rule of thumb, of choice, of the radiologists for the diagnosis of Pulmonary Embolism. Modified PIOPED criteria is described in detail in an article by Daniel F. Worsley and Abbas Alavi [1995]. The same software tool also incorporates the suggestions that were proposed by two nuclear medicine radiologists from the Medical College of Ohio, USA. The proposed Fuzzy Inference System is designed after the Mamdani model. The inputs required for the FIS, as determined using the PIOPED criteria, are as follows: (a) Number of segmental perfusions, (b) Number of non-segmental perfusions, (c) Ventilation/perfusion mismatch, (d) Chest x-ray abnormality, and (e) Pleural effusion. Two ventilation-perfusion scan features, namely number of segmental perfusions and number of non-segmental perfusions, were combined through a preprocessing procedure to generate a single input to the FIS for computational convenience. The rest of the inputs were directly provided to the FIS. Each input was assigned membership functions, which are defined by a set of curves. One of the most important considerations for the design of a FIS is the choice of membership functions and their ranges. The two radiologists provided valuable insight and recommendations in the development of membership functions and the fuzzy sets through their specialized medical expertise. Segmental and Non-Segmental Perfusions The Fuzzy Inference System needs to know the number of segmental perfusions (areas which are not getting perfused), which is then utilized to determine the degree to which the patient’s lungs are affected by Pulmonary Embolism. Non-segmental perfusions are those defects in the perfusion scan images, which do not cover the full size of a segment but nonetheless are important since they occupy a significant portion of the segment and this affects the proper perfusion of blood. The exact number of these defects are required. The fuzzy variable, Weight, has been represented using five membership functions to classify the inputs as presented in Figure 1. The fuzzy sets that are used for the fuzzy variable "Weight" are given by: ZE (Zero) – The sizes of all defects present in the V/Q scans after pre-processing that fall in the range of 0.0 to 2.0 have been deemed as normal and indicate no presence of perfusion defects. VLO (Very Low) – The sizes of all the defects on pre-processing that fall in the range between 2.0 to 4.0 indicate a low probabilty of Pulmonary Embolism. LO (Low) – The range for this fuzzy set is between 4.0 to 6.0. IM (Intermediate) – The range for this fuzzy set is between 6.0 to 8.5. HI (High) – This is the fuzzy set for defects of large sizes. The range for this fuzzy set starts at 8.5. The defects that fall under this set can be easily noticed occasionally covering the entire segment. The form of the membership function used for the fuzzy input “Weight” is Gaussian. Figure 1. Membership Functions for Fuzzy Input “Weight”. Weight Membership Function Plots Ventilation/Perfusion Mismatch By superimposing ventilation-perfusion scans, physicians try to assess the mismatch. It is this mismatch that is an important indication of Pulmonary Embolism. In case of large number of mismatches, the probability of Pulmonary Embolism increases. The fuzzy sets for the fuzzy variable "Vqdef", which is assigned a value in the range 0 to 100, are modeled using Gaussian functions and are defined as follows: CM (Correspondingly Matched) – the perfusion defect has a corresponding ventilation defect. MM (Multiple Matched) – each perfusion defect has a matching ventilation defect. SMM (Single Moderately Matched) – there is a slight degree of match between the images, perfusion and ventilation. UM (Unmatched) – perfusion and ventilation images show defects in different areas. Figure 2. Membership Functions for Fuzzy Input Vqdef. Chest x-ray Abnormality A chest X-ray will not be able to indicate conclusively if the patient suffers from Pulmonary Embolism, but it is the best way to check if the patient doesn’t suffer from it. On the presence of infiltrates (caused due to various other diseases), the radiologist can reduce the probability of Pulmonary Embolism being present in the patient and treat the patient with the proper procedure. The chest x-ray abnormality, "Cxrab", is obtained from the chest radiographs. The fuzzy sets that model “Cxrab”, Figure 3, are given by: NO (None) – Chest radiograph is normal. UML (Upper & Middle Lung) – some infiltrates present in the upper and middle lung area. LL (Lower Lung) – infiltrates present in the lower lung area. A Range of 0 to 100 has been determined to be appropriate for all possible values of the fuzzy variable “Cxrab”. The membership functions are also Gaussian in this case. Figure 3. Membership Functions for Fuzzy Input Cxrab. Membership Function Plots