(2001) 4, 56±62 ß 2001 Nature Publishing Group All rights reserved 1365±7852/01 $15.00 www.nature.com/pcan Correlation of transrectal ultrasound, computer analysis of transrectal ultrasound and histopathology of radical prostatectomy specimen HP Beerlage 1 *, RG Aarnink 1, ETh Ruijter 2, JA Witjes 1, H Wijkstra 1, CA van de Kaa 2, FMJ Debruyne 1 & JJMCH de la Rosette 1 1 Department of Urology, University Hospital Nijmegen, The Netherlands; and 2 Department of Pathology, University Hospital Nijmegen, The Netherlands A system for computerised analysis of ultrasonographic prostate images (AUDEX ˆ Automated Urologic Diagnostic EXpert system) for the detection of prostate carcinoma was developed. The ultimate goal is to develop a system that is reliable and non-observer dependent. Results of an earlier study with a small group were encouraging and this study describes the results of the computerised analysis in a larger group. Sixty-two patients who were scheduled to undergo a radical prostatectomy were prospectively analysed. The radical prostatectomy specimens were step-sectioned in the transverse plane, corresponding to the ultrasound pictures. Malignant regions identi ed by each study were quanti ed and compared by computer calculation. No correlation was observed between ultrasound analysis and pathology result. For the AUDEX analysis an overall sensitivity of 85% and a speci city of 18% with only a diagnostic accuracy of 57% was noticed when presence or absence of malignancy was evaluated by octant (total 496). When applying a cut-off value of 0.5 ml the numbers were 71%, 33% and 55%, respectively. Correlation was signi cantly better for the ventral octants. In this study the earlier results of our AUDEX system could not be con rmed. Although sensitivity was good, speci city and especially diagnostic accuracy were lower than expected. We have to conclude that the current settings are inappropriate for routine clinical use. (2001) 4, 56±62. Keywords: prostate carcinoma; ultrasound; computer interpretation Introduction Prostate carcinoma has become the leading cancer site in men in the United States and the second cause of death in men. 1 On the one hand, more efforts will have to be made to detect prostate carcinoma at an early stage to lower the mortality since extended disease cannot be cured. On the other, increased prostate awareness leaves us with a large number of patients who have an elevated prostatespeci c antigen (PSA) but have no abnormalities on *Correspondence: HP Beerlage MD FEBU, Urologist, University Hospital Nijmegen, PO BOX 9101, 6500 HB Nijmegen, The Netherlands. E-mail: H.Beerlage@uro.azn.nl Received 6 March 2000; accepted 1 November 2000 digital rectal examination (DRE) and normal transrectal ultrasound (TRUS). The sensitivity and speci city of TRUS alone is disappointingly low, 2,3 whereas the sensitivity of PSA is reasonably good but the speci city is poor, this especially resulting in a large number of unnecessary prostate biopsies. 4 Prostate biopsies may cause inconvenience for the patient and serious complications are associated with them. 5 DRE cannot be improved and PSA is the best prostate tumour marker available and none of the newer markers are clinically yet applicable. New imaging modalities should thus provide diagnostic information needed to reach higher sensitivity and speci city. In our department, a system for computerised analysis of ultrasonographic prostate images has been developed (AUDEX ˆ Automated Urologic Diagnostic EXpert system). Using the dependencies between 256 grey
tones, the computer is able to quantify texture and a number of parameters were calculated of local areas in each image. These parameters were correlated with the corresponding histology results, enabling the computer to differentiate between benign and malignant tissue. 6 This system was initially trained and evaluated with biopsies. 7 A preliminary study was published on the correlation between the AUDEX predictions and the pathology reports of 12 patients who underwent radical retropubic prostatectomy for localised prostate carcinoma. 8 In this small group a sensitivity of 75% and a speci city of 78% could be obtained for automated prostate cancer detection, comparable to the results obtained in the biopsy group, 81% and 77%, respectively. 6 We now report on a larger group in order to further assess the value of the AUDEX analysis. Patients and methods In this study 62 patients were prospectively analysed in a 4-year period. All patients had biopsy proven localised prostate carcinoma and subsequently underwent radical retropubic prostatectomy. A Kretz Combison 330 ultrasound scanner with a 7.5 Mhz multiplane transducer connected to a personal computer (486DX) with frame grabber (PC Vision Plus framegrabber, Difa Measurement Systems) was used to record a series of transrectal ultrasound images preoperatively. All examinations were performed by two urologists with great experience in TRUS studies. Transverse sections were imaged in a standardised fashion with the probe stabilised in a xture allowing to take 4 mm steps from the base of the prostate to the apex and resulting images were recorded by digitisation of the video signal. Following radical retropubic prostatectomy, the specimens were xed in formalin and step-sectioned at 4 mm in the transverse plane, thus corresponding to the ultrasound pictures. After paraf n embedding 5 mm, coupes were taken from each 4 mm section and stained with hematoxylin and eosin for histological examination. Photographs were taken of each 4 mm section and the malignant areas were outlined by the pathologists based on histopathological ndings. The recorded ultrasound images were interpreted by two ultrasound experts (HB and JR) and the areas suspicious for carcinoma were indicated on the image using manual delineation with the computer mouse. As criteria for suspicion asymmetry, capsular irregularity and hypoechogeneity were taken. The images were analysed twice: the rst analysis was based on ultrasound data alone while during the second analysis the clinical parameters and realtime ultrasound interpretation were known to the investigators as well. Interobserver variability could be judged in this way as well as the in uence of knowledge of clinical parameters on the interpretation of the ultrasound image. Automated image analysis was performed using the AUDEX system as described previously. 6,8 First order statistics use the grey scale value directly to describe the texture. Statistics of the second order use the dependencies in grey values that can be described in a so-called co-occurence matrix. 9 First, the prostate region was identi ed in the image. The image was then divided into windows of 30 6 45 pixels, re ecting image data of 0.5 6 0.5 cm. The settings in the construction of the co-occurrence matrix were: interpixel distance ˆ 3, angle ˆ 90. This means that the grey scale dependency was checked between pixels separated by three at four different angles 0, 90, 180 and 270. The resulting matrix was used to calculate the ve parameters that were most predictive in identifying prostate cancer with ultrasound data: signal to noise ratio, uniformity, contrast, inverse difference moment and entropy. The corresponding tissue was classi ed using the calculated parameters in the decision tree as constructed during the training of the system. A decision tree is a hierarchical description of the decisions made to partition the parameter space in hypercubes corresponding to the histology of the tissue. The construction was performed prospectively with cross-validation: building the decision tree with all samples but one and classifying the unused sample. By doing this for all samples, the tree with best predictions can be identi ed. 10 The results of this classi- cation were projected over the ultrasound image using a colour coding scale re ecting the probability of malignancy. The manually indicated areas of malignancy were projected separately on the ultrasound image. The result of each study was quanti ed by calculating the area of each malignant region. We used a 50% cut-off value in the probability for malignancy to calculate the areas of malignancy in the predicted images. As published previously, we interpreted the results locally by dividing the prostate in octants using the center point of the prostate. 8 For each octant the volume of malignant tissue indicated by the pathologist, urologist and the computer was calculated and stored in a database for comparison. Also the overall tumour volume for each technique was calculated. Results In Figure 1 a good example of a transverse ultrasound image with corresponding AUDEX and histological specimen cross section is shown. The computed probability for malignancy ranges from 0% (blue) to 100% (red). In Figure 2 however, an example is shown of a problematic AUDEX analysis. Due to calci cations the ventral part of the ultrasound image is too dark and is not calculated. In this case a tumour was present in the ventral part that was thus missed by AUDEX. The cut-off value used in this model is 50%, meaning that any region with a prediction greater than 50% was classi ed as malignant. Correlations between the different imaging techniques were calculated and compared with the gold standard being the histopathological examination. In the majority of cases (85%) tumours were found in the dorsal part of the prostate in this study; in 50% of cases a tumour was found ventrally, meaning that in 35% of cases tumour in both dorsal and ventral parts of the prostate was seen. No correlation between TRUS analysis, blind or with knowledge of data, and pathology result was seen. Subdivision of the prostate in left/right, in quarters or in octants did not disclose any area with a good correlation 57
58 Figure 1 (A) Prostate ultrasound image; (B) Corresponding AUDEX analysis (inner marked area suspicious for malignancy); (C) Corresponding pathology (dark area is malignant). between the several TRUS examinations and nal pathology. Second, the presence or absence of malignancy was evaluated for all octants. Whenever AUDEX or pathology indicated a cancerous area within one octant the entire octant was marked as malignant. Thus out of 62 patients a total of 496 octants were analysed. False and true positives and negatives are shown in Table 1. This results
59 Figure 2 (A) Prostate ultrasound image; (B) Corresponding AUDEX analysis (inner marked area suspicious for malignancy); (C) Corresponding pathology (dark area is malignant). in an overall sensitivity of 85% and a speci city of 18% only with a diagnostic accuracy of 57%. When applying a cut-off value of 0.5 ml the numbers are 71%, 33% and 55%, respectively (Table 2). In Table 3 the mean differences between cancer volume detected by AUDEX and nal pathology are shown using paired samples test. It is clear that AUDEX underestimates the tumour volume in all octants; this difference is
60 Table 1 octant Correlation of AUDEX analysis with histology for each ADL ADR AVL AVR BDL BDR BVL BVR SUM TP 40 41 26 22 42 36 20 18 245 FP 18 17 24 31 11 22 17 28 168 TN 0 0 6 4 3 0 15 9 37 FN 4 4 6 5 6 4 10 7 46 A ˆ apical (caudal), D ˆ dorsal, L ˆ left, TP ˆ true positive, TN ˆ true negative, B ˆ basal (cranial), V ˆ ventral, R ˆ right, FP ˆ false positive, FN ˆ false negative. Table 4 Method Calculated tumour volumes with different techniques Mean volume (ml) Pathology 4.72 AUDEX 2.13 TRUS 1 (blind) 1.29 TRUS 1 (data) 1.12 TRUS 2 (blind) 2.52 TRUS 2 (data) 1.55 TRUS 1 (blind) means interpretation of expert no. 1 without knowledge of clinical data and so forth. Table 2 Correlation of AUDEX analysis with histology for each octant for tumours > 0.5 ml ADL ADR AVL AVR BDL BDR BVL BVR SUM TP 33 33 23 20 34 28 20 16 207 FP 13 13 20 25 10 17 15 24 137 TN 5 4 10 10 4 5 17 13 68 FN 11 12 9 7 14 12 10 9 84 For explanation to abbreviations see footnote to Table 1. Table 3 Mean differences between cancer volume detected by AUDEX and nal pathology (paired samples test) Mean s.e.m. 95% CI Lower 95% CI Upper Signi cance ADL 70.97 0.43 71.83 70.11 0.027* ADR 70.99 0.38 71.76 70.22 0.012* AVL 70.13 0.41 70.96 70.70 0.754 AVR 70.17 0.36 0.90 70.55 0.629 BDL 71.57 0.45 72.49 70.67 0.001* BDR 71.38 0.43 72.26 70.51 0.002* BVL 70.66 0.37 71.41 0.09 0.083 BVR 70.56 0.43 71.42 0.31 0.202 * ˆ Signi cant (P < 0.05). For explanation to abbreviations see footnote to Table 1. signi cant for four of the eight octants: AUDEX signi cantly underestimates the cancer volume in all dorsal octants. Interpretation of ventral octants by AUDEX however does not differ signi cantly from the pathology results. So the correlation between AUDEX and nal pathology was especially poor in the dorsal part of the prostate. No correlation could be found with respect to tumour volume between the ultrasound ndings, computer ndings and nal pathology. Generally an underestimation of tumour volume was seen in all techniques when compared to pathology results as is demonstrated in Table 4. Knowledge of clinical data did not lead to a better outcome. Subdivision of the prostate in octants did not reveal certain areas in which the correlation was signi cantly stronger than in other areas. Discussion The AUDEX system uses texture analysis of the ultrasound images by means of co-occurence matrices as previously described. 6,9 This method is based on the statistical approach to texture which is concerned with the spatial distribution and dependence among the grey tones in a local area. In this series an analysis was made of 62 prostates in a prospective way comparing the results of ultrasound alone (two interpreters), ultrasound with clinical data (two interpreters) and AUDEX with the nal pathology result of the radical prostatectomy specimen being the gold standard. Although two experienced urologists analysed the images off-line, a large interobserver variability was noted and the correlation with pathology ndings was poor. Analysing the ultrasound images when clinical data were known did not improve the outcome signi cantly. A possible explanation however, may be the fact that the investigators knew that cancer had to be present in the prostate since all patients underwent radical prostatectomy. In addition, ultrasound is generally considered as a dynamic investigation, and interpretation from static images might not re ect the actual clinical situation. Nevertheless this is again a con rmation of the generally accepted perception that sensitivity and speci city of transrectal ultrasound are generally low and care should be taken to draw conclusions from it. This nding also stresses the importance of re nement and improvement of current imaging techniques. In this series sensitivity, speci city and diagnostic accuracy of AUDEX were found to be 85%, 18% and 57%, respectively. In our previous study with 12 patients however these numbers were 95%, 30% and 79%, respectively. 8 When the small volumes (< 10% malignant volume ratio or < 0.5 ml) were eliminated the numbers are 71%, 33% and 55% in the present study compared to 75%, 78% and 74% in our previous study. 8 In the present study the sensitivity data are comparable but speci city and particularly diagnostic accuracy however are apparently inferior. Indeed a diagnostic accuracy of 57% or 55% is low and the clinical value will be limited. In general some observations need to be made. In this study all images and data were entered into the computer and analysed quantitatively whereas in our previous study the volume of tumour on the photographs of the specimen as well as AUDEX images was estimated manually for each octant and divided into ve categories: 0%, 0 ± 10%, 10 ± 25%, 25 ± 50%, and 50 ± 100% malignant volume ratio. The absence of digital photographs of the pathology specimen did not allow exact quantitative analysis. Such an approach did not necessitate the selection of a cut-off value in prediction of 50%. This may be one of the explanations for the difference in diagnostic accuracy found.
In addition we noted that tumour volume was in general underestimated in all imaging modalities including AUDEX (Table 4). There are 256 different grey tones that can be distinguished with the AUDEX system, but not all grey tones were enumerated. Areas with an average grey tone < 35 and > 235 were considered to be artifacts and were not counted malignant. This means that tumours in these areas were not counted as such and this may (partially) explain the underestimation of the total tumour volume. This way of processing may also be responsible for the fact that the ventral tumours were signi cantly better detected than the dorsal tumours since artifacts (very dark areas) are more likely to occur at the dorsal part. We carefully assessed a number of possible explanations for the unsatisfactory results of the present series: (1) A change of patient characteristics over time such as age, PSA, tumour stage and grade. This was carefully evaluated within the group but no direct correlation could be found in this respect. The only difference found was the fact that in the previous series 76% of octants contained carcinoma whereas in the present series 59% of octants only contained carcinoma. (2) The original training set was based on biopsy data as was extensively described in our previous publications. 6,7 In spite of the fact that in our rst study the correlation between biopsy data and radical prostatectomy specimen was found to be quite good, 8 the inferior results of the present series may however still be related to the fact that the training set was based on biopsy data. (3) In our rst study on 12 patients the ratios were estimated whereas in the present group the tumours and tumour volumes were quantitatively assessed. (4) Drift of the ultrasound machine, due to changes in the crystal for instance, may cause inferior results. However in that case we would have expected a deterioration in time and careful analysis showed that there was no correlation with time whatsoever. This analysis excludes a universal drift in the machine; random drifts may however still be present, reducing the accuracy of the predictions. The results of this study prompted us not only to critically review our methods, as we did, but also to explore possibilities for improvement of the system. We believe that the principle of the texture analysis is appropriate for prostate cancer detection using ultrasound but it may be necessary to standardise the recording of the ultrasound image more extensively. Regular check-up of the ultrasound machine to verify the performance of the crystal and review of the settings is mandatory, although no structural in uence of these factors on the outcome could be demonstrated in this series. This does not mean that random drift might not be present and therefore setting review and adaptation remains necessary. Using phantoms and imaging them prior to each examination might reveal unexpected changes in the image properties. Furthermore corrections calculated from the phantom imaging will overcome problems associated with xed settings of the machine as currently required to allow prediction. Another factor that has to be taken into account is the fact that the original training of the AUDEX system was done using biopsy based data. First we could continue training the system using new biopsy data on a regular base, and in this way updating the training-set. Second, although in the previously published series of 12 radical prostatectomy specimens the results were quite good using the biopsy based data this may still account for the results of the present series. To overcome this problem we could take (a part of) the present series as a training set and analyse the individual cases using the new training set. A number of fascinating new developments in the eld of prostate ultrasound are currently under investigation such as 3-D Color Doppler and Power Doppler studies, contrast enhanced ultrasound and contrast angiosonography. 11 An incorporation of newer dynamic ultrasound parameters is a further way to possibly improve the system. These parameters describe dynamic processes such as blood ow. Carson et al already developed and clinically tested a model in which they incorporated dynamic parameters quantitatively, calculated from the 3-D Color Doppler ultrasound analysis of breast masses. 12 Conclusion In this prospective study on 62 patients the earlier results of our AUDEX system could not be con rmed. Although sensitivity was good, speci city and especially diagnostic accuracy were lower than expected. The tumour volume was generally underestimated. A number of possible causes for the deviating results are: lower number of tumour bearing octants, a quantitative versus semi-quantitative measurement, cut-off points of tumour volumes and grey tones and the fact that the original training set was based on biopsy data. We have to conclude that the current settings are inappropriate for routine clinical use. The dataset that was prospectively evaluated during this study might however serve as a starting point to teach the system with whole mount ultrasound sections. Additional information from blood ow studies can be added to the decision process. Acknowledgements The study was kindly supported by Kretz Technik AG, Zipf, Austria. References 1 Silverberg E, Lubera JA. Cancer statistics, 1989. CA Cancer J Clin 1989; 39: 3 ± 20. 2 Shinohara K, Scardino PT, Carter SS, Wheeler TM. Pathologic basis of the sonographic appearance of the normal and malignant prostate. Urol Clin North Am 1989; 16: 675 ± 691. 3 Aarnink RG et al. Transrectal ultrasound of the prostate: innovations and future applications. J Urol 1998; 159: 1568 ± 1579. 4 Montie JE, Meyers SE. De ning the ideal tumour marker for prostate cancer. Urol Clin North Am 1997; 24: 247 ± 252. 5 Beerlage HP, de Reijke TM, de la Rosette JJ. Considerations regarding prostate biopsies. Eur Urol 1998, 34: 303 ± 312. 61
62 6 Huynen AL et al. Analysis of ultrasonographic prostate images for the detection of prostatic carcinoma: the automated urologic diagnostic expert system. Ultrasound Med Biol 1994; 20: 1 ± 10. 7 De la Rosette JJ et al. Computerised analysis of transrectal ultrasonography images in the detection of prostate carcinoma. Br J Urol 1995; 75: 485 ± 491. 8 Giesen RJ et al. Computer analysis of transrectal ultrasound images of the prostate for the detection of carcinoma: a prospective study in radical prostatectomy specimens. J Urol 1995; 154: 1397 ± 1400. 9 Basset O, Sun Z, Mestas JL, Gimenez G. Texture analysis of ultrasonic images of the prostate by means of co-occurrence matrices. Ultrason Imaging 1993; 15: 218 ± 237. 10 Giesen RJ et al. Construction and application of hierarchical decision tree for classi cation of ultrasonographic prostate images. Med Biol Eng Comput 1996; 34: 105 ± 109. 11 Aarnink RG et al. Contrast angiosonography: a technology to improve Doppler ultrasound examinations of the prostate. Eur Urol 1999; 35: 9 ± 20. 12 Carson PL et al. 3-D Color Doppler image quanti cation of breast masses. Ultrasound Med Biol 1998; 24: 945 ± 952.