47. Diagnosis-Dynamic Characteristic Model

Toru Matsumoto, Akira Furukawa, Koji Suwa*, Nobuo Fukuda (*School of Dentistry at Tokyo, Nippon Dental University)

Keywords: image perception, decision-making model, search time, confidence rattng

A study was undertaken to construct a descriptive model for both the facilitation of an abnormality detection and the effect of search time on perception. In this study, we propose a model named the diagnosis-dynamic characteristic model (DDC) An outline of the model is as follows. Generally when medical doctors read an image and decide a "normal" or "abnormal" or "intermediate confi dence level", they require some time to search the image. In this model, it is supposed that the ttme required for searching an image depends on several factors of image quality, skill of the reader, facilita tion of diagnosis of an image and preconception before reading. Then psychological conflicts thinking "normal or abnormal" take place in the brain due to these factors, therefore, searching time is required. A mathematical formulation was made by incorporating these factors and it was fitted to the time(t) abnormal confidence level(p) data based on expertments.

To demonstrate the hypothesis derived from the DDC model, the model was applied to the oral CT image reading. A group of 20 oral CT images was carefully selected, which was comprised of 10 cases with abnormalities and TO cases of non-lesion. Thetr image quality were worsened by a thinned-out opera tion of pixels of the image matrix. Four groups (20 images/group) of different image quality which depended on the degree of the thinned-out operatton were generated in addition to the 20 originals. In the TOO CT images, 8 subjects searched for tumors. Eye movement data of subjects were measured by means of an eye tracking system. The length of time taken for each observation from beginning of image read ing until the decision-making, was calculated from the value of time accompanying eye-movement data. Moreover, after an image reading, each subject answered a confidence rating score of abnormality (with 4 steps) which was used to construct an ROC curve.

Results applied for the DDC model analysis for these experimental data showed that the model ex plains the relation between search time and Lconfidence level (Fig.14). That is, the confidence level (p) derived from the DDC model which exhibits the probability of abnormality varies non-linearly as the function of search time(t) of an image and the position of preconception which a subject had about a preconceived "normal" or "abnormal" at the start of diagnosis.

Finally, we confirmed that ordinal category scores of equal interval asked for the purpose of the ROC analysis (in Fig.14, they are 1, 2, 3 or 4) are equiva lent to integrate the continuous confidence level obtained from the DDC model for each observation of each subject. If search time measured at each obser vation and the .[)DC model is applied, although a judgement of subject is zero or one ("normal" or "abnormal), we can compose the ROC curve by ustng continuously-distributed probability of the DDC model.

Fig.14. A sample of the relationship between search time (t) and confidence rating (p) obtained by a fitting of the DDC model.