SRM benign & malignant, RCC classification, (clear cell RCC) grading literature review about ML(杂篇)

Benign & malignant：

Prediction of Benign and Malignant Solid Renal Masses: Machine Learning-Based CT Texture Analysis（Dec. 19）

土耳其伊斯坦布尔研究团队（全篇关键步骤工具包带过，研究流程具参考价值）

Dataset:

private; 79 patients with 84 solid renal masses (21 benign; 63 malignant) from a single center；恶性RCC亚型包括：透明细胞瘤，乳头状肾细胞癌和发色肾细胞癌；良型RCC亚型包含：肿瘤细胞瘤，脂肪少的血管平滑肌脂肪瘤  

Methods:

Results:

未增强的CT，具有良好可重复性的No. feature为198，对比增强CT，则为244。RF使用五个选定的增强CT纹理特征表现最佳。 曲线指标的acc和AUC分别为90.5％和0.915。消除了具有高度线性关系的特征后，acc和AUC分别略微增加至91.7％和0.916。

Methods details：

The extraction of texture features was performed using MaZda software which has a feature library of more than 300 features.

The reliability of texture features was evaluated with the intraclass correlation coefficient.

Feature selection was performed with the Waikato Environment for Knowledge Analysis toolkit version 3.8.2.(这里借助ATK完成，没读太懂)

The ML-based classifications were done using the Waikato Environment for Knowledge Analysis toolkit.（10 fold cv using AUC）

之后描述了对临床数据做的均值，中位数，正态分布，t检验，卡方检验等统计分析

Notes:

However, as a consequence of the significant similaritiesbetween malignant and benign renal masses in terms of imaging findings, reliable imaging criteria for the differentiation ofthese tumors do not exist and remain an important obstaclebefore treatment. Under such a condition where human analy-sis is unfruitful, ...

---

classification（分类；分型）：

现有的肾肿瘤影像AI研究焦点相对局限，还有很多亟待解决的临床问题没有涉及，如Bosniak 3级肾囊肿良恶性的影像鉴别和肾细胞癌常见亚型(透明细胞、乳头及嫌色肾细胞癌等) 的区分等（参考：Society of Abdominal Radiology disease-focused panel on renal cell carcinoma: update on past, current, and future goals（2018）； 人工智能在肾肿瘤影像中的应用（2018））。

 

受限于时间，但是目前手头文献收集结果基本说明，肾细胞癌的分类（分型）是比较新的领域，暂时不再考虑作为科研方向。

---

Grading：

Texture analysis and machine learning algorithms accurately predict histologic grade in small (< 4 cm) clear cell renal cell carcinomas: a pilot study（Dec. 19）

美国乔治华盛顿大学研究团队；流程具参考价值；To the best of our knowledge, this is the first paper using these methods to predict the histologic grade of small (< 4 cm) ccRCCs

Dataset:

private

Methods：

code援引：scikit learn

Results：

CM histogram skewness and GLRL short run emphasis (SRE) both had the highest AUCs of 0.82 in predicting ccRCC grade.

KNN, SVM, random forest and decision tree algorithms had the highest AUCs (0.97) predicting histologic grade using corticomedullary histogram features alone. These algorithms’ AUCs decreased using CM texture features alone (0.68–0.90) or combined CM histogram and texture features (0.78–0.93).

Things matters：

The hypervascular CM phase of ccRCC is most reflective of its histologic grade.

This is likely due to the integration of the large number of texture features that did not have significant differences between low- and high-grade ccRCCs.

The much higher discrimination in this study may be due to the small size of the ccRCCs; larger ccRCCs have higher propensity for internal hemorrhage, cystic or necrotic components, which introduces greater variability within histogram and texture features.

While RMB(renal mass biopsy) samples a very small portion of the tumor, texture analysis can “sample” a much larger por-tion of the mass with 2-dimensional cross-sectional image or three-dimensional volumetric data and may be able to better factor tumoral heterogeneity.

Sth have be demonstrated/well-recognized:

Correlation between ccRCC histologic grade and texture

CT histogram entropy was associated with histologic grade in RCCs greater than 7 cm

The strong correlation between increasing ccRCC size and higher Fuhrman histologic grade is well-recognized

Fuhrman histologic grading has some subjectivity in the assignment of nuclear pleomorphism

---

注：以上两篇文章的具体方法有一定代表性，整体流程是主流流程。