| Summary: | 碩士 === 國立聯合大學 === 電機工程學系碩士班 === 105 === Cancer is one of the top ten cause of death around the world. During a preliminary examination where there is a suspected tumor, a doctor will further confirm the existence of the tumor with immunohistochimal stained tissue section. A pathological examination can be used not just to classify tumors as benign or malignant but can also be utilized to confirm the invaded range, the appropriateness and integrity of the surgery, and the efficacy of the treatment. A pathological examination can also determine whether there is organ metastasis or lymphatic metastasis.
Several different staining methods can be used in immunohistochemical stained tissue sections. Not every method can be used for every organ. Ki-67 is a particular type of staining method that can be used for every organ; accordingly, this study is focused on ki-67 pathological sections. In clinical analysis, ki-67 is used to calculate the number of cells, but the current distinction is that ki-67 still relies on the doctor to manually calculate and circulate the approximate contour of a cell. Additionally, a survey has pointed out that in the next five to ten years, the number of pathologists will be considerably insufficient. This study seeks to develop a novel automatic cell counting algorithm for immunohistochemical stained tissue sections to reduce human resources and time, and to provide an objective reference.
There are two different ways to detect cells. One is object detection and the other is cell segmentation. Cell segmentation first identifies the cell contour and then uses the contour information to define the cell center. Using cell segmentation presents a potential problem of accurately segmentation the contour in both overlapping cells and in cells that are not uniform in color. If a cell is incorrectly segmented as multiple cells, the cell count will be inaccurate. In order to resolve this issue, this study uses the object detection method.
This study is divided into two parts. The first stage, which we will refer to as “seed detection,” detects the locations of the geometric cell center. This study is based on the single-pass voting method, which can change the estimated radius of cells, and then combined with the Gaussian mean shift to detect the position of the cell center. This study utilizes the gradient direction to check whether the estimated radius is much larger than the true cell radius. If the estimated radius is indeed significantly larger than the true cell radius, then the estimated radius should be decreased. The single-pass voting method is based on the Gaussian function. In order to use the Gaussian function, we must first determine the center position. It is unreasonable to predetermine the center position because we can’t be sure where the center position is located. We developed the second seed detection algorithm based on distance-weight in order to do so. The second stage is cell segmentation where the center point in previous stage is used to segment the nucleus region. After segmenting the nucleus region, the fuzzy C-means is applied to find the initial contour and adjust the contour with the Markov random field. In addition, this study has attempted to include ellipse fitting and Procrustes transformation. Using the shape information obtained through both ellipse fitting and Procrustes transformations as one of the potential functions in the Markov random field, we discovered that the result was poor.
After verifying the proposed method, the average F-score of the proposed method in seed detection was approximately 63%, and the average F-score of Qi’s method was approximately 55%. In case cell_04 and cell_18 the precision rates were approximately 80%, while in other cases, the precision rate was more than 60% .
In addition to the proposed method, we also conducted a pilot study for future research. In this pilot study, we use the FRST method as a pre-process, then performed seed detection. Comparing the results of the pilot study method and Qi’s method, we found the F-score of the pilot study method was approximately 49%. Although, the pilot study’s F-score didn’t reach 50%, it was still higher than that of Qi’s method. The F-score of Qi’s method was approximately 47%
|
|---|
