Abstract

Damrongsak Bulyalert
Division of Neurology, Department of Internal Medicine, Chiang Mai University, Faculty of Medicine, Thailand

When looking at a fine needle aspiration specimen of a breast mass, a pathologist is presented with cytological parameters such as clump thickness, uniformity of cell size, uniformity of cell shape, marginal adhesion, single epithelial cell size, bare nuclei, bland chromatin, normal nucleoli and mitoses. If malignancy is diagnosed, it is then confirmed using pathological sections obtained from the same breast mass.

A fundamental problem is whether these cytological parameters are weighed equally by the pathologist in order to reach such a diagnosis. Since all such parameters are presented at once under a microscope, there is no practical means for one to look at one parameter and completely disregard any others. Thus it is most likely that the diagnosis is reached by an overall impression and no single parameters could be used in isolation by the pathologist without being influenced by other parameters.

The purpose of this study was to demonstrate whether a pathologist might conceivably give some of these cytological parameters more weight than others in a diagnostic process. Using multilayer Perceptron (an artificial neural network, ANN) as a tool and breast cancer data from Wisconsin Breast Cancer Database, a public domain database available from University of California at Irvine, the ANN was trained to identify malignancy using these cytological parameters as inputs. For each set of cytological parameters an output was generated as either a benign or malignant lesion with pathological sections served as a gold standard. Experimental paradigms utilized single and multiple parameters, and omission of single and multiple parameters to train and test the ANN. A total of 683 cases were divided into 3 groups: 273 cases for training, 137 cases for cross validation and 273 cases for testing. Sensitivity and specificity of each paradigm was computed and then compared with others. Nemar’s test was used for statistical analysis.

The result from this study showed that some parameters such as single epithelial cell size and uniformity of cell size, when used as single inputs to the ANN, yielded better sensitivity and specificity than others while marginal adhesion, clump thickness and mitoses yielded poorer sensitivity and specificity. However, the best performance was obtained using all parameters. It also showed that omission of some parameters such as mitoses, cell thickness and marginal adhesion did not significantly affect the sensitivity and specificity of the ANN performance. This finding suggests that some parameters might conceivably receive more weight than others by a pathologist in the diagnostic process. It also suggests, and probably substantiates a long-held belief, that not all parameters are required to reach a correct diagnosis. This study discusses relative weights of various parameters using a quantitative approach, and the possibility that human decision making process might employ some specific strategies that normally could not be readily demonstrated.