We have presented two CAD systems that both emphasize an intelligible decision process for the prediction of breast cancer biopsy outcomes from BI-RADS input attributes. The first approach is based on decision-tree learning and the second on case-based reasoning using an entropic distance measure. We have evaluated and compared the performance of both systems using ROC analysis and bootstrap sampling. Both systems outperform the diagnosis decisions of the physicians. Hence, both systems have the potential to reduce the number of unnecessary breast biopsies. However, these results have been obtained using a retrospective approach using relatively small case databases containing only a few thousand cases. Therefore further investigations are needed, and these results probably need to be confirmed in a larger
prospective clinical study. A comparison of the ROC performances of the two proposed CAD approaches with a state of the art ANN approach shows that the CBR approach significantly outperforms the ANN approach, which in turn significantly outperforms the decision-tree approach. While there are significant but only small performance differences between the two proposed CAD approaches, there are fundamental differences regarding their properties. Hence, a decision to choose one over the other should be mainly based on evaluating these properties with regard to the intended clinical application. One fundamental difference is that while decision-tree systems are so-called eager learners, which induce a global model from the training data, their CBR counterparts are lazy learners, which directly use the training data to deduce a diagnosis suggestion. The advantage of a global decision-tree model is that it is a compact representation of the decision process, which can be understood simply by looking at the decision tree. Once the model is induced from the training data it can even be printed on a paper and applied to new cases without the need of a computer system. In contrast, the decision process of a CBR system is transparent to the physician in a different way. While it does not induce a global model, it induces local models for each new query case. These local models, which are simply based on the attributes and classification of the most similar database cases for a query case are also intelligible for the physician and can potentially model more complex decision processes than a global model.
เราได้นำเสนอสองระบบ CAD ทั้งเน้นกระบวนการตัดสินใจ intelligible สำหรับทำนายของผลตรวจชิ้นเนื้อมะเร็งเต้านมจาก BI RADS ใส่แอตทริบิวต์ วิธีแรกขึ้นอยู่กับการเรียนรู้ต้นไม้ตัดสินใจและที่สองในกรณีตามเหตุผลโดยใช้การวัดระยะทาง entropic เรามีประเมิน และเปรียบเทียบประสิทธิภาพของทั้งระบบที่ใช้วิเคราะห์ ROC และการเริ่มต้นระบบ ทั้งสองระบบมีประสิทธิภาพสูงกว่าการตัดสินใจการวินิจฉัยของแพทย์ที่ ดังนั้น ทั้งสองระบบมีศักยภาพในการลดจำนวนของประสาทการตรวจชิ้นเนื้อเต้านมไม่จำเป็น อย่างไรก็ตาม ผลลัพธ์เหล่านี้ได้ถูกรับใช้วิธีคาดที่ใช้ฐานข้อมูลกรณีที่ค่อนข้างเล็กประกอบด้วยเพียงไม่กี่พันกรณี ดังนั้น ต้องตรวจสอบเพิ่มเติม และผลลัพธ์เหล่านี้อาจจำเป็นต้องได้รับการยืนยันในขนาดใหญ่ prospective clinical study. A comparison of the ROC performances of the two proposed CAD approaches with a state of the art ANN approach shows that the CBR approach significantly outperforms the ANN approach, which in turn significantly outperforms the decision-tree approach. While there are significant but only small performance differences between the two proposed CAD approaches, there are fundamental differences regarding their properties. Hence, a decision to choose one over the other should be mainly based on evaluating these properties with regard to the intended clinical application. One fundamental difference is that while decision-tree systems are so-called eager learners, which induce a global model from the training data, their CBR counterparts are lazy learners, which directly use the training data to deduce a diagnosis suggestion. The advantage of a global decision-tree model is that it is a compact representation of the decision process, which can be understood simply by looking at the decision tree. Once the model is induced from the training data it can even be printed on a paper and applied to new cases without the need of a computer system. In contrast, the decision process of a CBR system is transparent to the physician in a different way. While it does not induce a global model, it induces local models for each new query case. These local models, which are simply based on the attributes and classification of the most similar database cases for a query case are also intelligible for the physician and can potentially model more complex decision processes than a global model.
การแปล กรุณารอสักครู่..