Application of intestinal CT texture analysis and nonlinear discriminant analysis in differential diagnosis of colorectal cancer and ulcerative colitis / 上海交通大学学报（医学版）

ABSTRACT

Objective • To evaluate the value of texture analysis in the discrimination of colorectal cancer (CRC) and ulcerative colitis (UC). Methods • The CT images of 61 CRC patients, 62 UC patients and 42 control objects were retrospectively analyzed. All the patients were pathologically proved and performed triphasic contrast-enhanced CT scan non-enhanced phase (NP), the arterial phase (AP) and the enteric phase (EP). The region of interest was drawn along the abnormal bowel wall's edge in each scan phase and texture features were generated by MaZda software. Based on 3 texture feature selection methods, the optimal subsets were generated and analyzed by 6 texture feature classification methods. The results were shown by misclassification rate (MCR). To compare the performance of texture-based classification and human visual classification, two radiologists with more than 10 years of gastrointestinal disease diagnostic experience analyzed the data. Results • The texture analysis based average MCR of differentiation between CRC and UC was (28.42±6.89)%, (28.19±4.07)%, (19.10±3.58)% in NP, AP, EP respectively. Compared with other texture feature classification methods, nonlinear discriminant analysis (NDA) was more accurate. In EP, NDA achieved an excellent classification result (MCR=12.61%). The average MCR between CRC and normally dilated bowel wall (NOR) was (13.33±7.21)%, (15.49±5.47)%, (6.74±3.02)%, while the average MCR between UC and NOR was (19.26±4.68)%, (20.04±6.63)%, (16.74±6.36)% in NP, AP and EP respectively. For visual classification between CRC and UC, the average MCR was (40.48±3.21)%, (35.71±1.60)%, (26.43±1.15)% in NP, AP, EP respectively. But the MCR of texture classification was lower than that of human vision classification, and computer texture classification had higher differential diagnosis rate. Conclusion • The CT-based texture analysis could be a feasible supplementary method to differentiate CRC from UC. NDA is more accurate than other classification methods, especially in EP.