We wanted to develop and test an artificial intelligence (AI) to assist physicians in making the thin-section CT diagnosis of diffuse pulmonary diseases.

The AI was composed of knowledge bases (KB) of 12 diffuse pulmonary diseases and an inference engine (IE). The KB of a disease included both the inclusion criteria (IC) and the exclusion criteria (EC), which were the clinical or thin-section CT findings that were known to be present or absent in that particular disease, respectively. From imputing the clinical or thin-section CT findings by the operator who was reading the thin-section CT, AI instantly executed the following two steps. First, the IE eliminated all diseases from the list which the EC had for those particular findings. Next, from a list of remaining diseases, the AI selected those diseases having those findings in its IC to formulate the 1st-step differential diagnosis (DD1). For the differential diagnosis in the next step, the reader could choose one more clinical or thin-section CT finding from the new list: [(all the findings in the IC or EC of DD1) - (the findings in the IC common to all the DD1s)]. The reader could proceed even further if needed. The system was tested on 10 radiology residents who solved 24 problems (two problems for each of 12 diffuse pulmonary diseases) without and then with the aid of the AI. The scores were compared using the Wilcoxon signed rank test.

An AI was made; it was composed of 280 rules (214 IC and 66 EC) and three interfaces (two for program management and another for problem solving). Contestants scored higher (p = 0.0078) using the AI (167 vs. 110 respectively), and they responded that they felt that the program was helpful in making decisions.

AI appeared to be helpful in making thin-section CT diagnosis.