The USC Andrew and Erna Viterbi School of Engineering USC Signal and Image Processing Institute USC Ming Hsieh Department of Electrical and Computer Engineering University of Southern California

Technical Report USC-EE-433

“A Study to Determine the Feasibility of Computer Diagnosis of the Pneumoconiosis”

by Richard P. Kruger

A study was undertaken to determine the feasibility for the possible automated mass diagnostic screening of pneumoconiosis radiographs. Two distinct textural feature extraction methods involving digital and coherent optical approaches were undertaken. The performance of the two automated diagnostic systems is described in detail in tables presented in the main body of this report. Analogous results are presented for diagnosis obtained from experienced radiologists asked to analyze the same films given to the automated systems.

As this was a feasibility study the available data base was necessarily limited. As the data base is expanded, the statistics of the measured features will become better known. Thus, one may conjecture that performance to date is most encouraging.

Normal/Abnormal classification accuracy (on a testing basis) for the optical system was 90.8%. The comparable percentages for the physicians ranged from 83.0% to 97.9%. It should be noted that the physician who's accuracy was 97.9% is a radiologist with extensive experience at diagnosing pneumoconiosis. Likewise, the missed abnormal testing rate for the optical system was 2.9%. The comparable rate for the physicians ranged from 1.0% to 6.9%.

The digital system also compared well with the physician results, though it is important to note that only a subset of the films was available for digital processing. For example, on a four class problem (a normal class and three progressively more severe abnormal classes), the accuracy of the digital results was essentially equivalent to the combined accuracy of the physicians.

To download the report in PDF format click here: USC-EE-433.pdf (4.0Mb)