【原】影像组学、深度学习代做服务-CT 扫描的 3D 图像分类

影像诊断小札记 2023-05-09 发布于江苏

展开全文

不依赖新机器、不依赖新的扫描技术做科研

科研案例：对体积数据进行医学图像分析的一种常用方法是使用深度 2D 卷积神经网络 (CNN)。这主要归因于 3D 数据的性质所带来的挑战：可变的体积大小、优化期间 GPU 耗尽。然而，在 2D CNN 中独立处理单个层面会故意丢弃深度信息，从而导致预期任务的性能不佳。因此，开发不仅可以克服繁重的内存和计算要求而且还可以利用 3D 信息的方法非常重要。为此，我们评估了一组体积均匀化方法来解决上述问题。该方法实现了CT扫描预测肺结核。

▲ 使用 3D CNN 处理 CT 扫描以进行结核病预测的统一技术

我们可以做些什么

我们依据上述方法，可以有很多二分类问题可以试验，也就对应有很多临床影像课题可以做，比如：

1.基于CT扫描3D图像鉴别细菌性肺炎与肺结核

2.基于CT扫描3D图像鉴别病毒性肺炎与细菌性肺炎

3.基于CT扫描3D图像鉴别肺泡性肺水肿与细菌性肺炎

4.其他（看各位的想象力）

不知道大家发现没有，这类课题有几个特征：1.解决的都是二分类问题；2.有3D数据最好，或者是多张图像，单张图像不适合该模型。

所以符合上述特征的，如果大家有自己想做的试验，可以私信我，你只需要准备数据，其他的交给我，费用方面只需要500，深度学习的结果就像炼丹，最后结果无法预测，所以前期小数据（30-50个）测试免费，当然最后如果结果还是不好，可以免费二次提交更多或者更换数据训练。

最终你将得到实验的结果

1.训练集以及测试集的准确率、AUC、假阳性率、假阴性率、真阳性率、真阴性率、敏感性、特异性：

Epoch 1/10070/70 - 43s - loss: 0.6815 - acc: 0.5929 - auc: 0.6465 - true_positives: 51.0000 - true_negatives: 32.0000 - false_positives: 38.0000 - false_negatives: 19.0000 - sensitivity_at_specificity: 0.7143 - specificity_at_sensitivity: 0.7143 - val_loss: 1.1329 - val_acc: 0.5000 - val_auc: 0.5000 - val_true_positives: 30.0000 - val_true_negatives: 0.0000e+00 - val_false_positives: 30.0000 - val_false_negatives: 0.0000e+00 - val_sensitivity_at_specificity: 0.0000e+00 - val_specificity_at_sensitivity: 0.0000e+00 - 43s/epoch - 609ms/stepEpoch 2/10070/70 - 26s - loss: 0.6817 - acc: 0.5571 - auc: 0.5819 - true_positives: 33.0000 - true_negatives: 45.0000 - false_positives: 25.0000 - false_negatives: 37.0000 - sensitivity_at_specificity: 0.6286 - specificity_at_sensitivity: 0.6000 - val_loss: 1.6583 - val_acc: 0.5000 - val_auc: 0.5000 - val_true_positives: 30.0000 - val_true_negatives: 0.0000e+00 - val_false_positives: 30.0000 - val_false_negatives: 0.0000e+00 - val_sensitivity_at_specificity: 0.0000e+00 - val_specificity_at_sensitivity: 0.0000e+00 - 26s/epoch - 365ms/stepEpoch 3/10070/70 - 25s - loss: 0.6142 - acc: 0.6071 - auc: 0.7063 - true_positives: 42.0000 - true_negatives: 43.0000 - false_positives: 27.0000 - false_negatives: 28.0000 - sensitivity_at_specificity: 0.7857 - specificity_at_sensitivity: 0.7429 - val_loss: 2.7207 - val_acc: 0.5000 - val_auc: 0.5000 - val_true_positives: 30.0000 - val_true_negatives: 0.0000e+00 - val_false_positives: 30.0000 - val_false_negatives: 0.0000e+00 - val_sensitivity_at_specificity: 0.0000e+00 - val_specificity_at_sensitivity: 0.0000e+00 - 25s/epoch - 364ms/stepEpoch 4/10070/70 - 25s - loss: 0.6133 - acc: 0.6429 - auc: 0.7172 - true_positives: 46.0000 - true_negatives: 44.0000 - false_positives: 26.0000 - false_negatives: 24.0000 - sensitivity_at_specificity: 0.8000 - specificity_at_sensitivity: 0.8143 - val_loss: 2.3975 - val_acc: 0.5000 - val_auc: 0.6189 - val_true_positives: 30.0000 - val_true_negatives: 0.0000e+00 - val_false_positives: 30.0000 - val_false_negatives: 0.0000e+00 - val_sensitivity_at_specificity: 0.1333 - val_specificity_at_sensitivity: 0.2667 - 25s/epoch - 359ms/stepEpoch 5/10070/70 - 25s - loss: 0.6181 - acc: 0.7000 - auc: 0.7263 - true_positives: 47.0000 - true_negatives: 51.0000 - false_positives: 19.0000 - false_negatives: 23.0000 - sensitivity_at_specificity: 0.7857 - specificity_at_sensitivity: 0.8571 - val_loss: 2.4028 - val_acc: 0.5000 - val_auc: 0.6550 - val_true_positives: 30.0000 - val_true_negatives: 0.0000e+00 - val_false_positives: 30.0000 - val_false_negatives: 0.0000e+00 - val_sensitivity_at_specificity: 0.1000 - val_specificity_at_sensitivity: 0.2333 - 25s/epoch - 361ms/stepEpoch 6/10070/70 - 26s - loss: 0.6041 - acc: 0.6929 - auc: 0.7554 - true_positives: 49.0000 - true_negatives: 48.0000 - false_positives: 22.0000 - false_negatives: 21.0000 - sensitivity_at_specificity: 0.8571 - specificity_at_sensitivity: 0.9143 - val_loss: 2.5357 - val_acc: 0.5000 - val_auc: 0.6261 - val_true_positives: 30.0000 - val_true_negatives: 0.0000e+00 - val_false_positives: 30.0000 - val_false_negatives: 0.0000e+00 - val_sensitivity_at_specificity: 0.5667 - val_specificity_at_sensitivity: 0.6333 - 26s/epoch - 366ms/stepEpoch 7/10070/70 - 26s - loss: 0.6330 - acc: 0.6143 - auc: 0.6857 - true_positives: 48.0000 - true_negatives: 38.0000 - false_positives: 32.0000 - false_negatives: 22.0000 - sensitivity_at_specificity: 0.7429 - specificity_at_sensitivity: 0.7429 - val_loss: 1.3485 - val_acc: 0.5000 - val_auc: 0.6994 - val_true_positives: 30.0000 - val_true_negatives: 0.0000e+00 - val_false_positives: 30.0000 - val_false_negatives: 0.0000e+00 - val_sensitivity_at_specificity: 0.6333 - val_specificity_at_sensitivity: 0.8000 - 26s/epoch - 378ms/stepEpoch 8/10070/70 - 26s - loss: 0.5817 - acc: 0.6786 - auc: 0.7423 - true_positives: 42.0000 - true_negatives: 53.0000 - false_positives: 17.0000 - false_negatives: 28.0000 - sensitivity_at_specificity: 0.7571 - specificity_at_sensitivity: 0.9143 - val_loss: 1.3542 - val_acc: 0.5000 - val_auc: 0.7022 - val_true_positives: 30.0000 - val_true_negatives: 0.0000e+00 - val_false_positives: 30.0000 - val_false_negatives: 0.0000e+00 - val_sensitivity_at_specificity: 0.7333 - val_specificity_at_sensitivity: 0.8000 - 26s/epoch - 367ms/stepEpoch 9/10070/70 - 26s - loss: 0.6388 - acc: 0.6286 - auc: 0.6648 - true_positives: 44.0000 - true_negatives: 44.0000 - false_positives: 26.0000 - false_negatives: 26.0000 - sensitivity_at_specificity: 0.7000 - specificity_at_sensitivity: 0.7429 - val_loss: 0.6765 - val_acc: 0.6000 - val_auc: 0.6933 - val_true_positives: 7.0000 - val_true_negatives: 29.0000 - val_false_positives: 1.0000 - val_false_negatives: 23.0000 - val_sensitivity_at_specificity: 0.7000 - val_specificity_at_sensitivity: 0.6667 - 26s/epoch - 369ms/stepEpoch 10/10070/70 - 26s - loss: 0.5788 - acc: 0.6929 - auc: 0.7685 - true_positives: 46.0000 - true_negatives: 51.0000 - false_positives: 19.0000 - false_negatives: 24.0000 - sensitivity_at_specificity: 0.8429 - specificity_at_sensitivity: 0.8429 - val_loss: 0.5777 - val_acc: 0.6500 - val_auc: 0.7439 - val_true_positives: 17.0000 - val_true_negatives: 22.0000 - val_false_positives: 8.0000 - val_false_negatives: 13.0000 - val_sensitivity_at_specificity: 0.7333 - val_specificity_at_sensitivity: 0.9000 - 26s/epoch - 368ms/step