Computer-aided diagnosis of odontogenic lesions

Computer-aided diagnosis of odontogenic lesions

Int. J. Oral Max illofac. Surg. 1986: 15: 592-596 (Key words: lesion. odontogenic: computer: diagnostic system) Computer-aided diagnosis of odontogen...

306KB Sizes 0 Downloads 63 Views

Int. J. Oral Max illofac. Surg. 1986: 15: 592-596 (Key words: lesion. odontogenic: computer: diagnostic system)

Computer-aided diagnosis of odontogenic lesions F. WIENER, D. LAUFER AND A. RIBAK

Department of Oral and Maxillofacial Surgery, Rambam Hospital. Faculty of Medicine, Technion Israellnstilute of Technology Haifa. Israel

4 cysts and 8 tumors of odontogenic origin have been described for the computer in terms of the prevalence of the associated clinical and radiological findings . Symptom prevalence as given in the literature and modified by our own experience was stated quantitatively as incidence ratios so that Bayes' formul a could be used to calculate the probability of each disorder for each set of patient data. The model was applied to 48 cases whose lesions were identified by histopathology following surgery. In all cases, the actu al lesion was listed in the computerproduced dilTerential diagnosis, in 94% of the eases the lesion headed the list, and in 75% a probability of 0.85 or greater was achieved. The model can thus be applied to patients awaiting surgery to obtain a reliable dilTerential diagnosis for guiding the medical stalT in the proper management of the patient.

ABSTRACT -

(Accepted for publication 20 October 1985)

In the past decade, there has been an increasing tendency to use the computer as an aid to medical practice. Thus we find computerized assessment of kidney disease", diabetic patients", and in the diseases of the digestive tract'. Programs have been developed for calculating the required composition of intravenous solutions', and for parenteral nutrition of newborns' and adult patients' under various degrees of malnutrition. In the area of oral and maxillofacial surgery, we recognize a series of cystic lesions and tumors of odontogenic origin. These

may differ from each other in their syrnptomology, location, their clinical and radiological aspect, or in their effect on the teeth and surrounding tissue. A computer program, taking into account the signs and symptoms typical of each of these disorders, can be of assistance to both trainees and house staff in determining the differential diagnosis with probabilistic accuracy. Since the treatment for most of these cases is surgery, it is very desirable that the preoperative diagnosis be as accurate as possible in order to properly prepare the patient and to perform the operation in an effective

COMPUTER IN DIAGNOSIS

manner. In this paper, we describe a model of the medical knowledge of odontogenic lesions which makes use of the findings and symptoms given in the literature correlated with anamnestic, clinical, and radiological findings in our own practice. This model is then applied to patient data to obtain the preoperative differential diagnosis and the results are compared to what was found at surgery.

Methods The decision rule for comparing patient data to the model of medical knowledge and calculating the probability that a given disease is consistent with the patient data is expressed by Bayes' formula:

593

Table l. Disorders considered in the model No. Disorders of cases Cysts I primordial cyst I to 2 dentigerous cyst 3 periapical (residual) cyst 26 4 odontogenic keratocyst o Tumors 5 ameloblastoma 6 6 calcifying epithelial odontogenic tumor 0 7 odontogenic myxoma I 8 peripheral cementing dysplasia - early 0 -late 0 9 central cementifying fibroma - early I -late 0 to ameloblastic fibroma o II odontoma 3 12 ameloblastic fibro-odontoma o total 48

P(D;)I1P(Sj ID i) j

P(D; I S)= LP(D;)I1P(Sj I D;) i

j

P(D;) is the probability that the patient has disease D; prior to considering any of the clinical findings. This could be the incidence of the disease in the given population or some other probabilistic ranking of the diseases. The probability that a patient with disease D, will have symptom S, is P( Sd Di} and, since the symptoms are assumed to be independent of eaeh other, the probability that the patient will have both symptoms SI and S2 is the product P(SdD i ) x P(S2 I D;). The probability that the patient with disease D, will have the entire symptom complex S, is P(D;)I1P(S} I DJ. The posterior (after having j

taken all the clinical findings into consideration) probability for disease Di; P( Dd Sf), is relative to all the other diseases (i.e., LP(D, I Sj)= I), so that I

the product in the numerator must be divided by the sum of this product for all the diseases in the model. The disorders to be diagnosed by the model included 4 cysts and 8 tumors of odontogenic origin. 2 of the tumors were separated into an early and late stage because of differences in their radiological finding with time. Table I lists the disorders considered in the model together with the number of patients with each disorder in the patient group on which we tested the model. Although 5 of the lesions had no representation in the patient group, we included them in the model

for completeness since they are fully described in the literature'r? and are considered an integral part of this area of medicine. Since our model was designed for the probabilistic classification of each case as one of the 12 disorders, all outcomes were assumed to be equally likely initially (i.e. P(Di) = I, for alii). The anamnestic, clinical and radiological findings associated with this disease area together with the prevalence of the relevant findings in each disorder was obtained from the literature?", The prevalence of findings are generally stated in qualitative terms, which for the purpose of the model had to be translated into incidence ratios P(S} IDi). Thus, for example, "majority" was taken to be 0.80, while "seldom" was assigned 0.05. For those lesions well represented in our patient group, some of the incidence ratios were modified to reflect our own experience. Thus while the literature reported the majority of patients with periapical cyst to be asymptomatic we found that the majority presented with swelling. This reflected the fact that the cases hospitalized in our department were with advanced pathology. The findings list, which also served as a questionnaire for patient data acquisition is given in Table 2. In order to test the validity of the disease model by showing that the incidence ratios assigned to each finding Icd to a unique description of the included diseases, a typical case was formulated for each disease. Each such typical case, when run against the medical knowledge model, obtained a

594

WIENER, LAUFER AND RIBAK

Table 2. Clinical findings for disease description Age Tooth child missing teenager unerupted young adult monvital adult impacted middle age displacement root Sex crown male female Symptoms asymp tomatic facial asymmetry swelling pain Location incisor, upper incisor, lower canine, upper canine, lower premolar, upper premolar, lower 'm olar, upper molar, lower 3rd molar, upper 3rd molar, lower

Balle luscent opaque opaque with luscent border bone displacement diffused border defined border Internal space calcified deposits honeycombed trabeculation toothlike smooth unilocul ar multilocul ar

probability of 0.85 or greater for the disease it was designed to illustrate. The model is thus capable of differentiating between each of the diseases considered. Data was obtained on 47 patients who were hosp italized and underwent surgery in our department in the last 4 years, one patient having 2 lesions, each in a different location. For each case, a histopathological diagnosis obtained after surgery was available. The data on each patient was entered into the computer and using Bayes' formula was compared to the model of medical knowledge to obtain the differential diagnosis. The computer report listed the clinical findings and the 4 most likely diseases ranked in descending order of probability. One such report is shown in Table 3.

Results The diagnostic capability of the model is summarized in Table 4. Of the 48 cases to which the model was applied, in 46 the computer's diagnosis based on the clinical and

radiological findings was identical with the histopathological conclusions obtained after surgery. In the other 2 cases, the histopathologically identified lesion appeared in the differential diagnosis list. As may be seen from the last column in Table 4, in 36 cases

Table 3. Computer generated ease report Patient 4 Clinical findings age: young adult sex: male symptoms: swelling symptoms: pain location: molar lower location: 3rd molar lower tooth: impacted tooth: root tooth: crown bone : luscent bone: defined border bone: bone displacement int. space: honeycombed trabeculation int. space: unilocul ar For these findings, the disease probabilities are: 0.861 5 ameloblastoma 0.112 7 odontogenic myxoma 0.015 10 ameloblastic fibroma 0.007 12 ameloblastic fibroodontoma

Table 4. Diagnostic capability of the model

cases correct diagnosis

Diseases* 2 3 5 7 9 II Total 10 26 6 I I 3 48 10 26 5 I 0 3 46

probability distribution: 0.951-1.000 0 8 16 I 0 0 3 28 0.901-0.950 0 I 3 I 0 0 0 5 0.851-0.900 I 0 I I 0 0 0 3 0.701-0.800 0 I I 0 0 0 0 2 0.601-0.700 0 0 I 0 0 0 0 I 0.501-0.600 0 0 I I 0 0 0 2 0.401-0.500 0 0 2 I I 0 0 4 0.301-0.400 0 0 I 0 0 0 0 I Mean probability (for each disease and tot al) 0.873, 0.955, 0.860, 0.656, 0.419, 0.064, 0.996, 0.838. * For disease name , see Table I.

COMPUTER IN DIAGNOSIS

the computer's diagnosis had a probability of 0.85 or more. Considering diagnoses with probabilities smaller than 0.85 as doubtful, we may state that the model has a diagnostic accuracy of 75% . As is also evident from Table 4, the model's diagnostic ability varied among the different diseases represented in our patient group. The mean probabilities shown in Table 4 for each disease and for the group as a whole include the 2 cases where the correct diagnosis appeared only in the differentiallist. The model was most successful in diagnosing odontoma and dentigerous cysts, but did more poorly for ameloblastoma. For periapical cysts, which had by far the largest representation in the patient group, the overall probability was reduced due to the inclusion of some doubtful cases. The diagnostic accuracy for periapical cyst was 77%, similar to the accuracy of the model in general. The diagnostic accuracy was 90% for dentigerous cysts, but only 50% for ameloblastoma.

Discussion Medical decision models can be roughly separated into 2 groups, those in which the decision rules are based on heuretics' or Boolean logic!'- 12 and those in which decisions are made on a statistical or Bayesian basisv', The first group is most appropriate for medical specialties where the decision making is a process which can be described in a flow chart. The computer system then simulates the physician expert's diagnostic or therapeutic reasoning. Although Bayesian decision making can also be used for sequential processes 10, it is best applied to areas such as odontogenic lesions in which all the data on the patient has already been recorded and the decision problem is essentially one of classification. The difficulty with the Bayesian model is that symptom prevalence must be stated

595

quantitatively as incidence ratios P(Sj I D;). In very few studies' has it been possible to obtain incidence ratios by actually observing a sufficiently large group of patients. The usual procedure is to do what we did in this study namely, to translate qualitative statements on symptom prevalence into quantitative estimates of incidence ratios. These " personal probabilities" are somewhat arbitrary, as for example 'majority' to which we assigned 0.80 can just as well be 0.70. If the diagnosis were to be based on a very small number of findings such a procedure may be disastrous. Fortunately, we have a list of 42 findings which must be considered, 8 or more of which will be relevant to a given patient. Bayes formula takes the effect of all findings into account simultaneously, thus mitigating inaccuracies in anyone of the findings. Moreover, it is not so much the magnitude of anyone incidence ratio that is essential in deciding between one disease and another, but rather the relati ve value. Thus, the presence of "pain" will make a diagnosis of dentigerous cyst (P(SID)=OAO) 4 times more likely than periapical cyst (P( S I D) = 0.10). The relative arbitrariness in using "personal probabilities"to formulate the medical knowledge model makes it essential that the model be applied to a group of patients with known diagnoses, such as we did in this study, and to readjust the "personal probabilities" as required. The fact that the model has a diagnotic accuracy of 75%, which is considered good', and that in all cases the true diagnosis appears in the differential list validates our approach. In 2 cases, the correct diagnosis was not at the top of the computer's differential diagnosis . One was an ameloblastoma for which the disease description is such that the disorder can be differentiated from the other diseases only if the patient presents the complete set of relevant findings (this also accounts for the reduced diagnostic ac-

596

WIENER, LAUFER AND RIBAK

curacy cited above for this disorder). The patient was asymptomatic and showed no bone displacement, whereas the disease description for ameloblastoma emphasized facial asymmetry and swelling as well as bone displacement. In the other, a case of central ccmentifying fibroma, the misdiagnosis was due to the fact that the patient had swelling instead of being asymptomatic and that the lesion was in the upper jaw instead of the lower. In this study, we have formulated and validated a model of odontogenic lesions which can be applied to patients awaiting oral surgery to obtain an accurate differential diagnosis. Such information can facilitate preparation of the patient for surgery and provide guidance to the surgeon in selecting the appropriate surgical procedure. The model may also prove useful as a tool in medical education and training in this area.

References 1. ARMON, Y., ROLL, D. & COLE, R.: The computer as an auxiliary tool in total parenteral nutrition of very low birth weight infants. Harefuali 1985: 108: 351-353. 2. DE Dm.IBAL, F. T., CLAMP, S. E., LEAPER, D. J., STANISLAND, J. R. & HORROCKS, J. c. Computer aided di agnosis of lower gastrointestinal tract disorders. Gastroenterology 1975: 68: 252-260. 3. EDWARDS, F. H.: Computer assisted planning of parenteral hyperalimentation therapy. c-u. Care Med. 1982: 10: 539-543.

4. FITZGERALD, L. T., OVERALL, J. E. & WILLIAMS, E. J.: A computer program for diagnosis of thyroid disease. Am. J. Roentgenol. 1966: 97: 901-905. 5. PAUKER, S. G. , GORRY, G. A., KASSlRER, J. P. & SCHWARTZ, W. 8.: Towards the simulation of medical cognition. Taking present illness by computer. Am. J . Med . 1976: 60: 981-996. 6. RICH, D. S., KARNOCK, C. M. & JEFFREY, L. P.: An evaluation of a micropcomputer in redicing the prepar ation time of parenteral nutrition solutions. J . Parent and Enteral. NlIlr. 1981: 6: 71-75. 7. SHAFER, W. G., HINE, M. K. & LEVY, B. Y.: Oral pathology, 3rd edition. Saunders, Philadelphia, 1974, pp. 236-282, 446-451. 8. THOMA, K. H. & GOLDMAN, H. M.: Oral pathology, 5th edition. Mosby, St. Louis 1960, pp. 780-809, 1168-1237. 9. TlECKE, W. R.: Oral pathology. McGraw-Hill, New York 1965, pp . 193-225. 10. WARNER, H. R., RUTIlERFORD, B. D. & HOUTCllENS, B.: A sequ ential Bayesean approa ch to history taking and d iagnosis. Comput. Biomed. Res. 1972: 5: 256-262. 11. WIENER, F. & WINAVER, J.: Computer simulation of medical reas oning in the a ssessment of kidney disease. Int. J. Biomed. Comput, 1977: 8: 203-215. 12. WIENER, E, FRENKEL, I., KANTER, Y. & BARZILAl, D.: Computerized medical decis ions in evaluating the diabetes patient. Comput. BioI. M ed. 1982: 12: 241-251. Address:

E Wiener Faculty of Medicine Technion - Israel Institute of Technology Efron Str eet Bat Galim, Haifa Israel