在物件偵測上，物體位置的標示經常以方框表示；然而在醫療影像中，有許多物件都有特定的幾何形狀，如圓形的瞳孔、橢圓的血球。在本研究中我們提出一種結構化區域提取網路的方法，以橢圓的錨點形狀偵測瞳孔、血球的輪廓，相比於方框的表示形式，物件輪廓的偵測方式會更接近於實體物件模樣。我們所提出的錨點形狀計算方法，能夠應用在許多現有的物件偵測網路架構中，實驗中，我們分別應用於兩個不同的偵測網路：Mask R-CNN與SSD，在其網路架構中，加入結構化的區域提取網路，分別對應瞳孔與血球偵測，用以直接估計出影像中的瞳孔、血球幾何定位參數。
實驗中，使用 Mask R-CNN 進行瞳孔定位的偵測結果，其 Dice Similarity Coefficient (DSC) 相似性係數達0.93，且在1230個瞳孔影像中，可正確偵測到1227個。而在使用 SSD 作血球偵測、計數實驗中，我們使用了公開的血球資料庫 ALL-IDB，其所有影像的血球個數約有 39000 顆，我們將資料分60%用於訓練、20%用於驗證，20%用於測試，偵測結果在 F1-Score 與 DSC 數據分別達到 97.45% 及 91.62%，並在部份重疊血球偵測上顯示出較方框精確的偵測結果，然而，在血球的旋轉角度定位上，平均誤差值約在25度左右，還有改善的空間。

In object detection, bounding boxes are often used as basic representations to localize objects. However, in medical images, many target objects have specific geometric shapes, such as circular pupils and elliptical blood cells. In this study, we propose a new structural region proposal network (RPN) that adopts parametric, elliptical anchor shapes to detect pupils and blood cells in medical images. Comparing with the adoption of bounding boxes as anchors in conventional RPN, the proposed elliptical anchor shapes and structural RPN may fit actual object shapes better. The proposed structural RPN can also be seamlessly integrated to many existing detection networks. In our experiments, the structural RPN is integrated into Mask R-CNN for pupil localization and SSD for blood cell detection to localize target objects and to estimate their shape parameters.
In the experiments for pupil location, the Dice similarity coefficient (DSC) reached 0.93 by using the Mask R-CNN with structural RPN. Among 1230 test pupil images, 1227 pupils are correctly detected. For blood cell detection and counting using SSD with structural RPN, a public blood cell database, ALL-IDB, which contains about 39,000 blood cells in images, is used. We use 60%, 20% and 20% of the data samples for training, validation and testing, respectively. The testing results of F1-score and DSC are 97.45% and 91.62%, respectively. The proposed approach using anchor shapes also has better qualitative performance in detecting occluded blood cells than the original SSD using bounding boxes. However, in estimating rotation angles of blood cells, the average estimation error of rotation is about 25 degrees, which leaves room for future improvement.

[1] J. Pearce, "The Marcus Gunn Pupil," Journal of Neurology, Neurosurgery & Psychiatry, vol. 61, no. 5, p. 520, 1996.
[2] C. David, "How to test for a relative afferent pupillary defect (RAPD)," Community Eye Health Journal, vol. 29, no. 96, pp. 68-69, 2016.
[3] A. A. Siddiqui, J. C. Clarke and A. Grzybowski, "William John Adie: the man behind the syndrome," Clinical & Experimental Ophthalmology, vol. 42, no. 8, pp. 778-784, 2014.
[4] D. G. F. Harriman and H. Garland, "The pathology of Adie's syndrome," Brain, vol. 91, no. 3, pp. 401-418, 1968.
[5] Y. M. Alomari, S. Norul, H. S. Abdullah, R. Z. Azma and K. Omar, "Automatic detection and quantification of WBCs and RBCs using iterative structured circle detection algorithm," 2014.
[6] P. Ghosh, D. Bhattacharjee, M. Nasipuri and K. D. Basu, "Automatic white blood cell measuring aid for medical diagnosis," in PACC, 2011.
[7] Z. He, T. Tan and Z. Sun, "Iris localization via pulling and pushing," in ICPR, 2006.
[8] C. A. Bastos, R. Tsang and G. D. Calvalcanti, "A combined pulling & pushing and active contour method for pupil segmentation," in ICASSP, 2010.
[9] E. T. Mahnaz, C. Lucas, S. Sadri and E. Y. K. Ng, "Analysis of breast thermography using fractal dimension to establish possible difference between malignant and benign patterns," Journal of Healthcare Engineering, vol. 1, no. 1, pp. 27-43, 2010.
[10] F. Fahmi, H. Marquering, G. Streekstra, L. Beenen, N. Janssen, C. Majoie and E. vanBavel, "Automatic Detection of CT Perfusion Datasets Unsuitable for Analysis due to Head Movement of Acute Ischemic Stroke Patients," Journal of Healthcare Engineering, vol. 5, no. 1, pp. 67-78, 2014.
[11] M. A. Abdullah, S. S. Dlay and W. L. Woo, "Fast and accurate pupil isolation based on morphology and active contour," in ICSIPA, 2014.
[12] S. Chen and J. Epps, "Efficient and robust pupil size and blink estimation from near-field video sequences for human–machine interaction," IEEE Trans. Cybernetics, vol. 44, no. 12, pp. 2356-2367, 2014.
[13] J. K. S. de Souza, M. A. da Silva Pinto, P. G. Vie, J. Baron and C. J. Tierra-Criollo, "An open-source, firewire camera-based, Labview-controlled image acquisition system for automated, dynamic pupillometry and blink detection," Computer Methods and Programs in Biomedicine, vol. 112, no. 3, pp. 607-623, 2013.
[14] M. Rizon, Y. Haniza, S. Puteh, A. Yeon, M. Shakaff, S. Abdul Rahman and M. Karthigayan, "Object detection using circular Hough transform," American Journal of Applied Sciences, vol. 2, no. 12, pp. 1606-1609, 2005.
[15] W. Gander, G. H. Golub and R. Strebel, "Least-squares fitting of circles and ellipses," BIT Numerical Mathematics, vol. 34, no. 4, pp. 558-578, 1994.
[16] M. Kass, A. Witkin and D. Terzopoulos, "Snakes: Active contour models," International Journal of Computer Vision, vol. 1, no. 4, pp. 321-331, 1988.
[17] T. F. Cootes, C. J. Taylor, D. H. Cooper and J. Graham, "Active shape models - their training and application," Computer Vision and Image Understanding, vol. 61, no. 1, pp. 38-59, 1995.
[18] S. A. Sahmoud and I. S. Abuhaiba, "Efficient iris segmentation method in unconstrained environments," Pattern Recognition, vol. 46, no. 12, pp. 3174-3185, 2013.
[19] T. L. Shen, B. I. Chuang, M. H. Shih and Y. N. Sun, "Fast pupil assessment for sensory evaluation from infrared video," in CVGIP, Taiwan, 2015.
[20] W. Fuhl, M. Tonsen, A. Bulling and E. Kasneci, "Pupil detection for head-mounted eye tracking in the wild: an evaluation of the state of the art," Machine Vision and Applications, vol. 27, no. 8, pp. 1275-1288, 2016.
[21] M. AbdulraheemFadhel , A. J. Humaidi and S. RazzaqOleiwi, "Image processing-based diagnosis of sickle cell anemia in erythrocytes," in NTICT, 2017.
[22] D. V. Hiren, "An automated blood cell segmentation using fuzzy based system," in IEEE Transactions on Medical Imaging, 2017.
[23] A. Bashir, Z. A. Mustafa, I. Abdelhameid and R. Ibrahem, "Detection of malaria parasites using digital image processing," in ICCCCEE), 2017.
[24] M. I. Razzak and S. Naz, "Microscopic blood smear segmentation and classification using deep contour aware CNN and extreme machine learning," in CVPRW, 2017.
[25] O. Ronneberger, P. Fischer and T. Brox, "U-net: Convolutional networks for biomedical image segmentation," in MICCAI, 2015.
[26] X. Li, H. Chen, X. Qi, Q. Dou, C. -W. Fu and P. -A. Heng, "H-denseunet: Hybrid densely connected unet for liver and tumor segmentation from CT volumes," IEEE Trans. Medical Imaging, vol. 37, no. 12, pp. 2663-2674, 2018.
[27] S. Ren, K. He, R. Girshick and J. Sun, "Faster R-CNN: Towards real-time object detection with region proposal networks," in NIPS, 2015.
[28] K. He, G. Gkioxari, P. Dollar and R. Girshick, "Mask R-CNN," in CVPR, 2017.
[29] R. Girshick, J. Donahue, T. Darrell and J. Malik, "Rich feature hierarchies for accurate object detection and semantic segmentation," in CVPR, 2014.
[30] R. Girshick, "Fast R-CNN," in ICCV, 2015.
[31] T. -Y. Lin, P. Dollar, R. Girshick, K. He, B. Hariharan and S. Belongie, "Feature pyramid networks for object detection," in CVPR, 2017.
[32] D. Pavllo, D. Grangier and M. Auli, "Quaternet: A quaternion-based recurrent model for human motion," in BMVC, 2018.
[33] E. Pervin and J. Webb, "Quaternions for computer vision and robotics," in CVPR, 1983.
[34] Z. Mirikharaji and G. Hamarneh, "Star shape prior in fully convolutional networks for skin lesion segmentation," in MICCAI, 2018.
[35] M. M. Moreno, I. González-Díaz and F. D. de María, "An elliptical shape-regularized convolutional neural network for skin lesion segmentation," in MICCAI, 2018.
[36] Y. Li, "Detecting Lesion Bounding Ellipses With Gaussian Proposal Networks," Baidu Research TR, 2019.
[37] A. Krizhevsky, I. Sutskever and G. Hinton, "Imagenet classification with deep convolutional neural networks," in NIPS, 2012.
[38] K. Simonyan and A. Zisserman, "Very deep convolutional networks for large-scale image recognition," in ICLR, 2015.
[39] K. He, X. Zhang, S. Ren and J. Sun, "Deep residual learning for image recognition," in CVPR, 2016.
[40] J. Long, E. Shelhamer and T. Darrell, "Fully convolutional networks for semantic segmentation," in CVPR, 2015.
[41] W. Liu, D. Anguelov, D. Erhan, C. Szegedy and S. E. Reed, "SSD: single shot multibox detector," CoRR, 2015.
[42] "Mask R-CNN source code," [Online]. Available: https://github.com/matterport/Mask_RCNN.
[43] "Non-Maximum-Suppression," [Online]. Available: https://medium.com/@chih.sheng.huang821/%E6%A9%9F%E5%99%A8-%E6%B7%B1%E5%BA%A6%E5%AD%B8%E7%BF%92-%E7%89%A9%E4%BB%B6%E5%81%B5%E6%B8%AC-non-maximum-suppression-nms-aa70c45adffa.
[44] "ROI-ALIGN," [Online]. Available: https://blog.csdn.net/xiamentingtao/article/details/78598511.
[45] L. R. Dice, "Measures of the amount of ecologic association between species," Ecology, vol. 26, no. 3, pp. 297-302, 1945.
[46] T. L. Shen, B. I. Chuang, M. H. Shih and Y. N. Sun, "Real-Time Pupil Segmentation and Assessment System Using Appearance-based Circle Matching in High Resolution Infrared Eye Image Sequences," Technical Report of NCKU, 2017.