Diagnostic Aid Based on Computer Extracted Features

Jiang et al. have performed another observer study to assess the clinical potential of computer classification of clustered microcalcifications as malignant or benign [59]. Their technique consisted of eight computer-extracted features and an ANN classifier. These features are listed in Table 1 and have been described earlier in this chapter. The computer aid to radiologists was an estimate of the likelihood that a microcalcification cluster represented a malignancy. This study was the first to evaluate the clinical potential of computer classification based on computer-extracted features that is completely independent from radiologists' interpretation of mammograms. Like the Getty study, this study also demonstrated that the classification aid can improve radiologists' accuracy in diagnosing breast lesions as malignant or benign. In addition, this study also indicated that computer classification can help radiologists make better biopsy recommendations.

Ten radiologists who had experience in reading mammograms served as observers. Five observers were attending radiologists who read mammograms as part of clinical practice but who did not read mammograms exclusively. These radiologists had read mammograms for an average of 9 years at the time of the experiment. They represented a sample of radiologists who were responsible for interpreting the majority of mammograms in the United States. Expert mammographers, who are few in number, and radiologists who are not qualified to read mammograms according to MQSAwere not sampled. The other five observers were senior radiology residents. The residents represented a sample of radiologists who had just started interpreting mammograms with little previous experience except that obtained during residency training.

They used mammograms from 104 patients in whom a suspicious cluster of microcalcifications was found. In all cases, the diagnosis was confirmed by either surgical or percutaneous core biopsies. Of the 104 cases, 46 were malignant and 58 were benign. The cancer cases consisted of 9 invasive cancers and 37 cases of ductal carcinoma in situ (DCIS). In each case, the original standard-view and magnification-view mammograms were available for the observers to interpret. However, the computer analysis of the mammograms was done on the standard views only. The set of cases was derived from a consecutive biopsy series, leaving out only cases that contained masses and those that had missing films. The proportion of cancers was enriched for this series [79].

Figure 8 shows the results in terms of ROC analysis comparing the average performance of the radiologists with and without the computer aid. The observers' unaided ROC curve showed the poorest performance, with an Az value of 0.61. The observers' aided or CAD ROC curve showed a substantially improved performance, with an Az of 0.75. Statistical tests confirmed that the improvement in the Az value was statistically significant (p < 0.0001). These results show that the radiologists were able to use the computer classification to modify their confidence regarding lesion malignancy and they substantially improved their diagnostic accuracy. The computer-only ROC curve, with an Az value of 0.80, indicates better performance than the observers' CAD performance, suggesting that although the radiologists were able to benefit from the computer aid, they were unable to use it optimally. It is possible that combining their own confidence with the computer aid, both of which were continuous quantities, was not a trivial task for a human observer. In other words, the maximum potential in the gain of diagnostic accuracy was not realized in the study.

Radiologists' ability to recommend biopsy for cancers and not to recommend biopsy for benign lesions was also improved by using the computer aid. Without the computer aid, the average sensitivity of the 10 observers in terms of making biopsy recommendations was 73.5% and their average specificity was 31.6%. When the computer classification was provided as an aid, the average sensitivity was improved to 87.4% and the average specificity was improved to 41.9%. Use of the computer classification resulted in a simultaneous increase both in sensitivity (14%) and in specificity (10%). When compared in terms of averages across the 10 observers, each observer recommended biopsy for 6.4 additional malignant cases when they used the computer aid (p = 0.0006), and each observer recommended follow-up for 6.0 additional benign lesions when they used the computer aid (p = 0.003). As a result of these changes, the hypothetical positive biopsy yield — the fraction of cases recommended for biopsy that are malignant — increased on average from 46% to 55%.

Two example cases are shown in Figs 9 and 10 to illustrate

FIGURE 9 Mammograms of (a) standard CC view and (b) magnification MIO view showing (4 inch x 3 inch) a cluster of microcalcifications in the lower-inner quadrant of the left breast. The histological diagnosis was invasive ductal carcinoma. The computer-estimated likelihood of malignancy was 60%. The computer result helped two radiologists correctly diagnose this case (see Table 2). Reprinted from [59] with permission.

FIGURE 9 Mammograms of (a) standard CC view and (b) magnification MIO view showing (4 inch x 3 inch) a cluster of microcalcifications in the lower-inner quadrant of the left breast. The histological diagnosis was invasive ductal carcinoma. The computer-estimated likelihood of malignancy was 60%. The computer result helped two radiologists correctly diagnose this case (see Table 2). Reprinted from [59] with permission.

FIGURE 10 Mammograms of (a) standard MLO view and (b) magnification LM view showing (4 inch x 3 inch) a cluster of microcalcifications in the upper-outer quadrant of the right breast. The histological diagnosis was fibrocystic changes, fibroadenoma, and adenosis. The computer-estimated likelihood of malignancy was 13%. The computer aid helped all five radiologists avoid recommending biopsy for this case (see Table 3). Reprinted from [59] with permission.

FIGURE 10 Mammograms of (a) standard MLO view and (b) magnification LM view showing (4 inch x 3 inch) a cluster of microcalcifications in the upper-outer quadrant of the right breast. The histological diagnosis was fibrocystic changes, fibroadenoma, and adenosis. The computer-estimated likelihood of malignancy was 13%. The computer aid helped all five radiologists avoid recommending biopsy for this case (see Table 3). Reprinted from [59] with permission.

the effect of computer classification on radiologists' diagnostic confidence and on their biopsy recommendations. Figure 9 shows an invasive ductal carcinoma seen in the lower-inner quadrant of the left breast. Figure 10 shows a benign lesion consisting of fibrocystic changes, fibroadenoma, and adenosis, seen in the upper-outer quadrant of the right breast. The histological diagnoses of both cases were obtained through surgical biopsy. The computer results reported a 60% likelihood of malignancy for the malignant case (Fig. 9), and 13% for the benign case (Fig. 10). For the malignant case, all five attending radiologists reported higher levels of suspicion with the computer aid, and two radiologists changed their clinical recommendations from follow-up to surgical or alternative tissue sampling (Table 2). Consequently, the computer aid helped two radiologists correctly diagnose an additional cancer that they initially had missed. For the benign case, all five attending radiologists reported lower levels of suspicion with the computer aid, and all five radiologists changed their clinical recommendation from biopsy to follow-up (Table 3). In this manner, the computer aid helped all five radiologists avoid recommending biopsy for a benign lesion.

10 Ways To Fight Off Cancer

10 Ways To Fight Off Cancer

Learning About 10 Ways Fight Off Cancer Can Have Amazing Benefits For Your Life The Best Tips On How To Keep This Killer At Bay Discovering that you or a loved one has cancer can be utterly terrifying. All the same, once you comprehend the causes of cancer and learn how to reverse those causes, you or your loved one may have more than a fighting chance of beating out cancer.

Get My Free Ebook


Post a comment