We investigate deep learning-based quantum readout error mitigation techniques. Mitigation of readout errors is one major task to reliably execute quantum algorithms. There is a gap between already known quantum algorithms and their application in science and technology. Current classical error mitigation methods mostly have high costs or low accuracy, while quantum methods require extra gates or qubits that are not available. We propose a deep learning-based framework to solve this task with high quality and low resource requirements in the online phase. Our work gives a thorough analysis of the presented models, which exceeds by far the validation of their pure capability. Different aspects towards a fully scalable model are investigated and successively implemented. We find efficient models to mitigate measurements with up to 9 qubits and offer advantages for up to 11 qubits. The framework is publicly available and can be applied and extended to any neural network model architecture. Our findings aim to guide the research in directions, where fully scalable error mitigation methods can be developed.