资源受限的循环U-Net: 在生物医学图像分割中的高性能、低资源消耗的方法

21420资源受限分割的循环U-Net0魏旺�于凯成*乔希姆∙休戈诺特帕斯卡尔∙富亚马修∙萨尔兹曼CVLab，EPFL，1015洛桑0{first.last}@epfl.ch0摘要0最先进的分割方法依赖于非常深的网络，这些网络不总是容易在没有非常大的训练数据集的情况下进行训练，并且在标准GPU上运行相对较慢。在本文中，我们引入了一种新颖的循环U-Net架构，它保持了原始U-Net[33]的紧凑性，同时显著提高了其性能，以至于在几个基准测试中超过了最先进的方法。我们将展示其在手部分割、视网膜血管分割和道路分割等多个任务中的有效性。我们还引入了一个大规模的手部分割数据集。01. 引言0尽管最近的语义分割方法取得了令人印象深刻的结果[6, 17,18,46]，但它们需要非常深的网络，其架构往往专注于高分辨率和大规模数据集，并依赖于预训练的主干网络。例如，最先进的模型，如Deeplab [5, 6]，PSPnet [46]和Re�neNet[17]，使用ResNet101[15]作为它们的主干网络。这导致GPU内存使用量和推理时间较高，并使它们在功率受限的环境中不太理想，尽管实时性能仍然是必需的，例如使用增强现实头戴设备的内部资源进行手部分割。这一问题已经通过ICNet[45]等架构得到解决，但代价是性能显著下降。更重要的是，训练非常深的网络通常需要大量的训练数据或接近ImageNet[10]的图像统计数据，这在生物医学图像分割等领域可能不合适，其中更紧凑的U-Net架构仍然占主导地位[33]。在本文中，我们认为这些最先进的方法不适用于资源受限的情况，并引入了一种新颖的循环U-Net架构，它保持了原始U-Net [33]的紧凑性，同时显著提高了其性能，达到了0�平等贡献。0（a）每秒帧数（FPS）0mIoU0            我们的0                 步骤1  循环简单0                   循环中间0步骤10气泡大小    #参数（百万）  ∝0<1M0约8M0约40M约118M0真实值 T=1 T=2 T=3 图像0图1：速度与准确性。每个圆圈代表模型在我们的键盘手部数据集上使用TitanX（Pascal）GPU的每秒帧数和mIoU准确性的性能。每个圆圈的半径表示模型的参数数量。对于我们的循环方法，我们在1、2和3次迭代后绘制这些数字，并在底部行显示相应的分割结果。我们的方法的性能以红色绘制，其他缩写在第4.2节中定义。ICNet[45]比我们稍快，但代价是显著的准确性下降，而Re�neNet[17]和DeepLab[6]在该数据集上都较慢且准确性较低，可能是因为没有足够的训练样本来学习它们的许多参数。0在5个手部分割数据集上，我们的模型优于当前最先进的模型，其中一个数据集在图1中展示，并且还有一个视网膜血管分割数据集。我们的模型仅有0.3百万参数，比基于ResNet101的DeepLabv3+ [6]和Re�neNet[17]的模型分别少40和118百万个权重。这有助于解释为什么我们可以在专门的任务上胜过最先进的网络：预训练的ImageNet特征不一定是最好的，训练集也没有CityScapes[9]那么大。因此，大型网络往往会过拟合，并且性能不如从头开始训练的紧凑模型。标准的U-Net将图像作为输入，处理它，并直接返回输出。相比之下，我们的循环CNNs0CNNCNNCNN Conv Convht−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.AAACxnicjVHLSsNAFD2Nr1pfVZdugkVwVRIp6LLgpsuK9gG1lGQ6rUPzIjNRShH8Abf6aeIf6F94Z0xBLaITkpw5954zc+/1k0BI5TivBWtpeWV1rbhe2tjc2t4p7+61ZZyljLdYHMRp1/ckD0TEW0qogHeTlHuhH/COPznX8c4tT6WIoys1TXg/9MaRGAnmKaIu5UANyhWn6phlLwI3BxXkqxmXX3CNIWIwZAjBEUERDuBB0tODCwcJcX3MiEsJCRPnuEeJtBllccrwiJ3Qd0y7Xs5GtNee0qgZnRLQm5LSxhFpYspLCevTbBPPjLNmf/OeGU99tyn9/dwrJFbhhti/dPPM/+p0LQojnJkaBNWUGEZXx3KXzHRF39z+UpUih4Q4jYcUTwkzo5z32TYaaWrXvfVM/M1kalbvWZ6b4V3fkgbs/hznImifVF3CF7VKvZaPuogDHOKY5nmKOhpookXeYzziCc9Ww4qszLr7TLUKuWYf35b18AGdDZBRAAACxnicjVHLSsNAFD2Nr1pfVZdugkVwVRIp6LLgpsuK9gG1lGQ6rUPzIjNRShH8Abf6aeIf6F94Z0xBLaITkpw5954zc+/1k0BI5TivBWtpeWV1rbhe2tjc2t4p7+61ZZyljLdYHMRp1/ckD0TEW0qogHeTlHuhH/COPznX8c4tT6WIoys1TXg/9MaRGAnmKaIu5UANyhWn6phlLwI3BxXkqxmXX3CNIWIwZAjBEUERDuBB0tODCwcJcX3MiEsJCRPnuEeJtBllccrwiJ3Qd0y7Xs5GtNee0qgZnRLQm5LSxhFpYspLCevTbBPPjLNmf/OeGU99tyn9/dwrJFbhhti/dPPM/+p0LQojnJkaBNWUGEZXx3KXzHRF39z+UpUih4Q4jYcUTwkzo5z32TYaaWrXvfVM/M1kalbvWZ6b4V3fkgbs/hznImifVF3CF7VKvZaPuogDHOKY5nmKOhpookXeYzziCc9Ww4qszLr7TLUKuWYf35b18AGdDZBRAAACxnicjVHLSsNAFD2Nr1pfVZdugkVwVRIp6LLgpsuK9gG1lGQ6rUPzIjNRShH8Abf6aeIf6F94Z0xBLaITkpw5954zc+/1k0BI5TivBWtpeWV1rbhe2tjc2t4p7+61ZZyljLdYHMRp1/ckD0TEW0qogHeTlHuhH/COPznX8c4tT6WIoys1TXg/9MaRGAnmKaIu5UANyhWn6phlLwI3BxXkqxmXX3CNIWIwZAjBEUERDuBB0tODCwcJcX3MiEsJCRPnuEeJtBllccrwiJ3Qd0y7Xs5GtNee0qgZnRLQm5LSxhFpYspLCevTbBPPjLNmf/OeGU99tyn9/dwrJFbhhti/dPPM/+p0LQojnJkaBNWUGEZXx3KXzHRF39z+UpUih4Q4jYcUTwkzo5z32TYaaWrXvfVM/M1kalbvWZ6b4V3fkgbs/hznImifVF3CF7VKvZaPuogDHOKY5nmKOhpookXeYzziCc9Ww4qszLr7TLUKuWYf35b18AGdDZBRAAACxnicjVHLSsNAFD2Nr1pfVZdugkVwVRIp6LLgpsuK9gG1lGQ6rUPzIjNRShH8Abf6aeIf6F94Z0xBLaITkpw5954zc+/1k0BI5TivBWtpeWV1rbhe2tjc2t4p7+61ZZyljLdYHMRp1/ckD0TEW0qogHeTlHuhH/COPznX8c4tT6WIoys1TXg/9MaRGAnmKaIu5UANyhWn6phlLwI3BxXkqxmXX3CNIWIwZAjBEUERDuBB0tODCwcJcX3MiEsJCRPnuEeJtBllccrwiJ3Qd0y7Xs5GtNee0qgZnRLQm5LSxhFpYspLCevTbBPPjLNmf/OeGU99tyn9/dwrJFbhhti/dPPM/+p0LQojnJkaBNWUGEZXx3KXzHRF39z+UpUih4Q4jYcUTwkzo5z32TYaaWrXvfVM/M1kalbvWZ6b4V3fkgbs/hznImifVF3CF7VKvZaPuogDHOKY5nmKOhpookXeYzziCc9Ww4qszLr7TLUKuWYf35b18AGdDZBRsAAACxHicjVHLSsNAFD2Nr1pfVZdugkVwVRIp6LIgiMsW7ANqkWQ6raHTJMxMhFL0B9zqt4l/oH/hnTEFtYhOSHLm3HvOzL03TEWktOe9Fpyl5ZXVteJ6aWNza3unvLvXVkkmGW+xRCSyGwaKiyjmLR1pwbup5MEkFLwTjs9NvHPHpYqS+EpPU96fBKM4GkYs0EQ11U254lU9u9xF4Oeggnw1kvILrjFAAoYME3DE0IQFAih6evDhISWujxlxklBk4xz3KJE2oyxOGQGxY/qOaNfL2Zj2xlNZNaNTBL2SlC6OSJNQniRsTnNtPLPOhv3Ne2Y9zd2m9A9zrwmxGrfE/qWbZ/5XZ2rRGOLM1hBRTallTHUsd8lsV8zN3S9VaXJIiTN4QHFJmFnlvM+u1Shbu+ltYONvNtOwZs/y3Azv5pY0YP/nOBdB+6TqE27WKvVaPuoiDnCIY5rnKeq4RAMt6/2IJzw7F45wlJN9pjqFXLOPb8t5+ABbCI9qAAACxHicjVHLSsNAFD2Nr1pfVZdugkVwVRIp6LIgiMsW7ANqkWQ6raHTJMxMhFL0B9zqt4l/oH/hnTEFtYhOSHLm3HvOzL03TEWktOe9Fpyl5ZXVteJ6aWNza3unvLvXVkkmGW+xRCSyGwaKiyjmLR1pwbup5MEkFLwTjs9NvHPHpYqS+EpPU96fBKM4GkYs0EQ11U254lU9u9xF4Oeggnw1kvILrjFAAoYME3DE0IQFAih6evDhISWujxlxklBk4xz3KJE2oyxOGQGxY/qOaNfL2Zj2xlNZNaNTBL2SlC6OSJNQniRsTnNtPLPOhv3Ne2Y9zd2m9A9zrwmxGrfE/qWbZ/5XZ2rRGOLM1hBRTallTHUsd8lsV8zN3S9VaXJIiTN4QHFJmFnlvM+u1Shbu+ltYONvNtOwZs/y3Azv5pY0YP/nOBdB+6TqE27WKvVaPuoiDnCIY5rnKeq4RAMt6/2IJzw7F45wlJN9pjqFXLOPb8t5+ABbCI9qAAACxHicjVHLSsNAFD2Nr1pfVZdugkVwVRIp6LIgiMsW7ANqkWQ6raHTJMxMhFL0B9zqt4l/oH/hnTEFtYhOSHLm3HvOzL03TEWktOe9Fpyl5ZXVteJ6aWNza3unvLvXVkkmGW+xRCSyGwaKiyjmLR1pwbup5MEkFLwTjs9NvHPHpYqS+EpPU96fBKM4GkYs0EQ11U254lU9u9xF4Oeggnw1kvILrjFAAoYME3DE0IQFAih6evDhISWujxlxklBk4xz3KJE2oyxOGQGxY/qOaNfL2Zj2xlNZNaNTBL2SlC6OSJNQniRsTnNtPLPOhv3Ne2Y9zd2m9A9zrwmxGrfE/qWbZ/5XZ2rRGOLM1hBRTallTHUsd8lsV8zN3S9VaXJIiTN4QHFJmFnlvM+u1Shbu+ltYONvNtOwZs/y3Azv5pY0YP/nOBdB+6TqE27WKvVaPuoiDnCIY5rnKeq4RAMt6/2IJzw7F45wlJN9pjqFXLOPb8t5+ABbCI9qAAACxHicjVHLSsNAFD2Nr1pfVZdugkVwVRIp6LIgiMsW7ANqkWQ6raHTJMxMhFL0B9zqt4l/oH/hnTEFtYhOSHLm3HvOzL03TEWktOe9Fpyl5ZXVteJ6aWNza3unvLvXVkkmGW+xRCSyGwaKiyjmLR1pwbup5MEkFLwTjs9NvHPHpYqS+EpPU96fBKM4GkYs0EQ11U254lU9u9xF4Oeggnw1kvILrjFAAoYME3DE0IQFAih6evDhISWujxlxklBk4xz3KJE2oyxOGQGxY/qOaNfL2Zj2xlNZNaNTBL2SlC6OSJNQniRsTnNtPLPOhv3Ne2Y9zd2m9A9zrwmxGrfE/qWbZ/5XZ2rRGOLM1hBRTallTHUsd8lsV8zN3S9VaXJIiTN4QHFJmFnlvM+u1Shbu+ltYONvNtOwZs/y3Azv5pY0YP/nOBdB+6TqE27WKvVaPuoiDnCIY5rnKeq4RAMt6/2IJzw7F45wlJN9pjqFXLOPb8t5+ABbCI9qs2AAACxnicjVHLSsNAFD2Nr1pfVZdugkVwVRIp6LLgpsuK9gFaSjKdtqF5MZkopQj+gFv9NPEP9C+8M05BLaITkpw5954zc+/10zDIpOO8Fqyl5ZXVteJ6aWNza3unvLvXzpJcMN5iSZiIru9lPAxi3pKBDHk3FdyL/JB3/Mm5induuciCJL6S05T3Im8UB8OAeZKoy3Hf6ZcrTtXRy14ErgEVmNVMyi+4wQAJGHJE4IghCYfwkNFzDRcOUuJ6mBEnCAU6znGPEmlzyuKU4RE7oe+IdteGjWmvPDOtZnRKSK8gpY0j0iSUJwir02wdz7WzYn/znmlPdbcp/X3jFRErMSb2L9088786VYvEEGe6hoBqSjWjqmPGJdddUTe3v1QlySElTuEBxQVhppXzPttak+naVW89HX/TmYpVe2Zyc7yrW9KA3Z/jXATtk6pL+KJWqdfMqIs4wCGOaZ6nqKOBJlrkPcIjnvBsNazYyq27z1SrYDT7+Lashw/hSJACAAACxnicjVHLSsNAFD2Nr1pfVZdugkVwVRIp6LLgpsuK9gFaSjKdtqF5MZkopQj+gFv9NPEP9C+8M05BLaITkpw5954zc+/10zDIpOO8Fqyl5ZXVteJ6aWNza3unvLvXzpJcMN5iSZiIru9lPAxi3pKBDHk3FdyL/JB3/Mm5induuciCJL6S05T3Im8UB8OAeZKoy3Hf6ZcrTtXRy14ErgEVmNVMyi+4wQAJGHJE4IghCYfwkNFzDRcOUuJ6mBEnCAU6znGPEmlzyuKU4RE7oe+IdteGjWmvPDOtZnRKSK8gpY0j0iSUJwir02wdz7WzYn/znmlPdbcp/X3jFRErMSb2L9088786VYvEEGe6hoBqSjWjqmPGJdddUTe3v1QlySElTuEBxQVhppXzPttak+naVW89HX/TmYpVe2Zyc7yrW9KA3Z/jXATtk6pL+KJWqdfMqIs4wCGOaZ6nqKOBJlrkPcIjnvBsNazYyq27z1SrYDT7+Lashw/hSJACAAACxnicjVHLSsNAFD2Nr1pfVZdugkVwVRIp6LLgpsuK9gFaSjKdtqF5MZkopQj+gFv9NPEP9C+8M05BLaITkpw5954zc+/10zDIpOO8Fqyl5ZXVteJ6aWNza3unvLvXzpJcMN5iSZiIru9lPAxi3pKBDHk3FdyL/JB3/Mm5induuciCJL6S05T3Im8UB8OAeZKoy3Hf6ZcrTtXRy14ErgEVmNVMyi+4wQAJGHJE4IghCYfwkNFzDRcOUuJ6mBEnCAU6znGPEmlzyuKU4RE7oe+IdteGjWmvPDOtZnRKSK8gpY0j0iSUJwir02wdz7WzYn/znmlPdbcp/X3jFRErMSb2L9088786VYvEEGe6hoBqSjWjqmPGJdddUTe3v1QlySElTuEBxQVhppXzPttak+naVW89HX/TmYpVe2Zyc7yrW9KA3Z/jXATtk6pL+KJWqdfMqIs4wCGOaZ6nqKOBJlrkPcIjnvBsNazYyq27z1SrYDT7+Lashw/hSJACAAACxnicjVHLSsNAFD2Nr1pfVZdugkVwVRIp6LLgpsuK9gFaSjKdtqF5MZkopQj+gFv9NPEP9C+8M05BLaITkpw5954zc+/10zDIpOO8Fqyl5ZXVteJ6aWNza3unvLvXzpJcMN5iSZiIru9lPAxi3pKBDHk3FdyL/JB3/Mm5induuciCJL6S05T3Im8UB8OAeZKoy3Hf6ZcrTtXRy14ErgEVmNVMyi+4wQAJGHJE4IghCYfwkNFzDRcOUuJ6mBEnCAU6znGPEmlzyuKU4RE7oe+IdteGjWmvPDOtZnRKSK8gpY0j0iSUJwir02wdz7WzYn/znmlPdbcp/X3jFRErMSb2L9088786VYvEEGe6hoBqSjWjqmPGJdddUTe3v1QlySElTuEBxQVhppXzPttak+naVW89HX/TmYpVe2Zyc7yrW9KA3Z/jXATtk6pL+KJWqdfMqIs4wCGOaZ6nqKOBJlrkPcIjnvBsNazYyq27z1SrYDT7+Lashw/hSJACAAACxnicjVHLSsNAFD2Nr1pfVZdugkVwVRIp6LLgpsuK9gFaSpJO26F5MZkopQj+gFv9NPEP9C+8M05BLaITkpw5954zc+/105Bn0nFeC9bS8srqWnG9tLG5tb1T3t1rZ0kuAtYKkjARXd/LWMhj1pJchqybCuZFfsg6/uRcxTu3TGQ8ia/kNGW9yBvFfMgDTxJ1Oe67/XLFqTp62YvANaACs5pJ+QU3GCBBgBwRGGJIwiE8ZPRcw4WDlLgeZsQJQlzHGe5RIm1OWYwyPGIn9B3R7tqwMe2VZ6bVAZ0S0itIaeOINAnlCcLqNFvHc+2s2N+8Z9pT3W1Kf994RcRKjIn9SzfP/K9O1SIxxJmugVNNqWZUdYFxyXVX1M3tL1VJckiJU3hAcUE40Mp5n22tyXTtqreejr/pTMWqfWByc7yrW9KA3Z/jXATtk6pL+KJWqdfMqIs4wCGOaZ6nqKOBJlrkPcIjnvBsNazYyq27z1SrYDT7+Lashw/jqJADAAACxnicjVHLSsNAFD2Nr1pfVZdugkVwVRIp6LLgpsuK9gFaSpJO26F5MZkopQj+gFv9NPEP9C+8M05BLaITkpw5954zc+/105Bn0nFeC9bS8srqWnG9tLG5tb1T3t1rZ0kuAtYKkjARXd/LWMhj1pJchqybCuZFfsg6/uRcxTu3TGQ8ia/kNGW9yBvFfMgDTxJ1Oe67/XLFqTp62YvANaACs5pJ+QU3GCBBgBwRGGJIwiE8ZPRcw4WDlLgeZsQJQlzHGe5RIm1OWYwyPGIn9B3R7tqwMe2VZ6bVAZ0S0itIaeOINAnlCcLqNFvHc+2s2N+8Z9pT3W1Kf994RcRKjIn9SzfP/K9O1SIxxJmugVNNqWZUdYFxyXVX1M3tL1VJckiJU3hAcUE40Mp5n22tyXTtqreejr/pTMWqfWByc7yrW9KA3Z/jXATtk6pL+KJWqdfMqIs4wCGOaZ6nqKOBJlrkPcIjnvBsNazYyq27z1SrYDT7+Lashw/jqJADAAACxnicjVHLSsNAFD2Nr1pfVZdugkVwVRIp6LLgpsuK9gFaSpJO26F5MZkopQj+gFv9NPEP9C+8M05BLaITkpw5954zc+/105Bn0nFeC9bS8srqWnG9tLG5tb1T3t1rZ0kuAtYKkjARXd/LWMhj1pJchqybCuZFfsg6/uRcxTu3TGQ8ia/kNGW9yBvFfMgDTxJ1Oe67/XLFqTp62YvANaACs5pJ+QU3GCBBgBwRGGJIwiE8ZPRcw4WDlLgeZsQJQlzHGe5RIm1OWYwyPGIn9B3R7tqwMe2VZ6bVAZ0S0itIaeOINAnlCcLqNFvHc+2s2N+8Z9pT3W1Kf994RcRKjIn9SzfP/K9O1SIxxJmugVNNqWZUdYFxyXVX1M3tL1VJckiJU3hAcUE40Mp5n22tyXTtqreejr/pTMWqfWByc7yrW9KA3Z/jXATtk6pL+KJWqdfMqIs4wCGOaZ6nqKOBJlrkPcIjnvBsNazYyq27z1SrYDT7+Lashw/jqJADAAACxnicjVHLSsNAFD2Nr1pfVZdugkVwVRIp6LLgpsuK9gFaSpJO26F5MZkopQj+gFv9NPEP9C+8M05BLaITkpw5954zc+/105Bn0nFeC9bS8srqWnG9tLG5tb1T3t1rZ0kuAtYKkjARXd/LWMhj1pJchqybCuZFfsg6/uRcxTu3TGQ8ia/kNGW9yBvFfMgDTxJ1Oe67/XLFqTp62YvANaACs5pJ+QU3GCBBgBwRGGJIwiE8ZPRcw4WDlLgeZsQJQlzHGe5RIm1OWYwyPGIn9B3R7tqwMe2VZ6bVAZ0S0itIaeOINAnlCcLqNFvHc+2s2N+8Z9pT3W1Kf994RcRKjIn9SzfP/K9O1SIxxJmugVNNqWZU